A Future That Works Full Report MGI January 2017

A FUTURE THAT WORKS:
AUTOMATION, EMPLOYMENT,
AND PRODUCTIVITY
JANUARY 2017
Since its founding in 1990, the McKinsey Global Institute (MGI) has sought
to develop a deeper understanding of the evolving global economy. As the
business and economics research arm of McKinsey & Company, MGI aims
to provide leaders in the commercial, public, and social sectors with the facts
and insights on which to base management and policy decisions. The Lauder
Institute at the University of Pennsylvania ranked MGI the worlds number-one
private-sector think tank in its 2015 Global Think Tank Index.
MGI research combines the disciplines of economics and management,
employing the analytical tools of economics with the insights of business
leaders. Our micro-to-macro methodology examines microeconomic
industry trends to better understand the broad macroeconomic forces
affecting business strategy and public policy. MGIs in-depth reports have
covered more than 20 countries and 30 industries. Current research focuses
on six themes: productivity and growth, natural resources, labor markets, the
evolution of global financial markets, the economic impact of technology and
innovation, and urbanization.
Recent reports have assessed the economic benefits of tackling gender
inequality, a new era of global competition, Chinese innovation, and
digital globalization. MGI is led by four McKinsey & Company senior
partners: JacquesBughin, JamesManyika, JonathanWoetzel, and
FrankMattern, MGIs chairman. MichaelChui, SusanLund, AnuMadgavkar,
SreeRamaswamy, and JaanaRemes serve as MGI partners. Project
teams are led by the MGI partners and a group of senior fellows and include
consultants from McKinsey offices around the world. These teams draw
on McKinseys global network of partners and industry and management
experts. Input is provided by the MGI Council, which coleads projects
and provides guidance; members are AndresCadena, RichardDobbs,
KatyGeorge, RajatGupta, EricHazan, EricLabaye, AchaLeke, ScottNyquist,
GaryPinkus, ShirishSankhe, OliverTonby, and EckartWindhagen. In addition,
leading economists, including Nobel laureates, act as research advisers.
The partners of McKinsey fund MGIs research; it is not commissioned by any
business, government, or other institution. For further information about MGI
and to download reports, please visit www.mckinsey.com/mgi.
Copyright McKinsey & Company 2017
A FUTURE THAT WORKS:

AUTOMATION, EMPLOYMENT,
AND PRODUCTIVITY
JANUARY 2017
James Manyika | San Francisco

Michael Chui | San Francisco
Mehdi Miremadi | Chicago
Jacques Bughin | Brussels
Katy George | New Jersey
Paul Willmott | London
Martin Dewhurst | London
PREFACE
Automation is an idea that has inspired science fiction writers and futurologists for more than
a century. Today it is no longer fiction, as companies increasingly use robots on production
lines or algorithms to optimize their logistics, manage inventory, and carry out other core
business functions. Technological advances are creating a new automation age in which
ever-smarter and more flexible machines will be deployed on an ever-larger scale in the
workplace. In reality, the process of automating tasks done by humans has been under way
for centuries. What has perhaps changed is the pace and scope of what can be automated.
It is a prospect that raises more questions than it answers. How will automation transform
the workplace? What will be the implications for employment? And what is likely to be its
impact both on productivity in the global economy and on employment?
This report was produced as part of the McKinsey Global Institutes overall research on
the impact of technology on business and society, and specifically our ongoing research
program on the future of work and the potential impacts on the global economy of data and
analytics, automation, robotics and artificial intelligence. In this report, we analyze how a
wide range of technologies could potentially automate current work activities that people are
paid to do in the global workforce, and what the impact could be on global productivity. This
work does not define what the new activities and occupations that will be developed will
be, nor does it analyze in depth how the economic gains of automation will be distributed or
provide specific policy recommendations for governments. While we consider a broad range
of automation technologies, we do not focus specifically on any particular technologies.
We realize that this area of research is evolving rapidly, given the pace of technological
advancement, and we plan to update the perspectives presented in this report regularly.
The research was led by JamesManyika, a director of the McKinsey Global Institute
and McKinsey senior partner based in San Francisco; MichaelChui, an MGI partner in
San Francisco; and MehdiMiremadi, a McKinsey partner based in Chicago. MGI and
McKinsey senior partner JacquesBughin and McKinsey senior partners MartinDewhurst,
KatyGeorge, AndrewGrant, BillSchaninger, StefanSpang, and PaulWillmott guided
and contributed to the research. We would especially like to acknowledge the work
of SeanKane, RickCavolo, and TongChen, who each headed the research team at
different times over the course of the two-year project and who played invaluable roles
in coordinating and driving it forward. The team comprised AnujAbrol, JaredBarnett,
JacksonBeard, LilyCheng, JoshCogan, ArielleCopeland, SamDoniger, RachelGarber,
PaulGilson, BobGlied, SartajGrover, GauriGupta, BenjaminHarrison, AlexHinch,
TanayJuipuria, DanielLanger, KunalMehta, AndreyMironenko, VaishalPatel, StevenPecht,
JonathanSands, SanthoshSuresh, AdamTourgee, RoshinUnnikrishnan, JeanXin,
RogerYang, GordonYu, and VickiYu.
We are grateful to colleagues within McKinsey who provided valuable advice and analytical
support: JonathanAblett, PraveenAdhi, AdrianaAragon, MandarPAtre, GretchenBerlin,
DilipBhattacharjee, UrsBinggeli, BedeBroome, AndersBrun, JamieChang,
ZinaidaCherevan, AriChester, JeffreyCondon, SebasCordero, SubbiReddyDwarampudi,
AlanFitzgerald, ChristopherForr-Rydgren, DanielGarnier, PancoGeorgiev, TimoGlave,
DavideGronchi, SaifHameed, HolgerHurtgen, AndrewJordan, JoshuaKatz, RichardKelly,
AimeeKim, SajalKohli, PaulKolter, MichelleLin, JoyLong, MurdockMartin, AnneMartinelli,
IngaMaurer, JanMischke, SebastianMontero, ThomasNetzer, RobinNuttall,

CatherineOConnor, DarrylPiasecki, LaurenRatner, RonRitter, KelseyRobinson,
ElidoRodriguez, ArushiSaksena, SamSamdani, RichardSellschop, JeongminSeong,
RamanSharma, VivienSinger, KatarzynaSmietana, GiacomoSonnino, JonathanTilley,
SoyokoUmeno, SariVarpa, SusanneWagner, RichardWard, RobWhiteman, TingWu, and
MauricioZamora.
This report was edited and produced by MGI senior editor PeterGumbel, editorial
production manager JuliePhilpot, senior graphic designers MarisaCarder,
MargoShimasaki, and PatrickWhite, and data visualization editor RichardJohnson.
RebecaRobboy, MGI director of external communications, managed communications.
We would like to thank our academic advisers for their invaluable insights and guidance:
MartinBaily, Bernard L. Schwartz Chair in Economic Policy Development and senior
fellow and director of the Business and Public Policy Initiative at the Brookings Institution;
MatthewJ.Slaughter, Paul Danos Dean of the Tuck School and the Earl C. Daum 1924
Professor of International Business at Dartmouth College; and HalVarian, chief economist
at Google.
Many other external experts informed our research, including those whom we surveyed
about the timing of technological progression. We are deeply grateful for their insight and
assistance. In addition, our research has benefited from the work of other researchers,
including DaronAcemoglu, DavidH.Autor, ErikBrynjolfsson, CarlBenediktFrey,
JasonFurman, AndrewMcAfee, MichaelOsborne, and TimOReilly.
This report contributes to MGIs mission to help business and policy leaders understand
the forces transforming the global economy, identify strategic locations, and prepare for the
next wave of growth. As with all MGI research, this work is independent and has not been
commissioned or sponsored in any way by any business, government, or other institution.
While we are grateful for all the input we have received, the report is ours, including any
errors. We welcome your comments on this research at MGI@mckinsey.com.
Jacques Bughin
Director, McKinsey Global Institute
Senior Partner, McKinsey & Company
Brussels
James Manyika
San Francisco
Jonathan Woetzel
Shanghai
January 2017
A restaurant in Yiwu, Zhejiang province of China.

Visual China Group/Getty Images
CONTENTS
In brief
HIGHLIGHTS
23
Executive summary Page 1

Overview of select recent studies on the impact of automation and
futureofwork Page 21
Technology is advancing
rapidly
1. The new frontier Page 23

53
Automation technologies including robotics and artificial intelligence have advanced rapidly,
but some key hurdles still need to be cleared
2. The technical potential forautomation Page 29
A vision of future automation
Almost half of the work activities in all sectors across the economy have the potential to be
automated by adapting currently demonstrated technologies
114
3. Five case studies Page 53
How automation could potentially transform hospital emergency departments, aircraft
maintenance, mortgage brokering, oil and gas control rooms, and grocery stores
Humans will interact more
with machines
4. Factors affecting the pace and extent of automation Page 65

Five technical and economic factors that will together determine the pace at which
automation arrives in the workplace
5. An engine of productivity Page 87
Automation could compensate for demographic trends by giving a major boost to the
global economy
6. Preparing for disruption Page 109
The implications of automation for business leaders, policy makers, and workers
Technical appendix Page 119
Bibliography Page 133
IN BRIEF
A FUTURE THAT WORKS: AUTOMATION,

EMPLOYMENT, AND PRODUCTIVITY
Advances in robotics, artificial intelligence, and machine
learning are ushering in a new age of automation, as
machines match or outperform human performance in a
range of work activities, including ones requiring cognitive
capabilities. In this report, part of our ongoing research
into the future of work, we analyze the automation
potential of the global economy, the factors that will
determine the pace and extent of workplace adoption,
and the economic impact associated with its potential.
Automation of activities can enable businesses
to improve performance, by reducing errors and
improving quality and speed, and in some cases
achieving outcomes that go beyond human
capabilities. Automation also contributes to
productivity, as it has done historically. At a time
of lackluster productivity growth, this would give a
needed boost to economic growth and prosperity
and help offset the impact of a declining share of the
working-age population in many countries. Based
on our scenario modeling, we estimate automation
could raise productivity growth globally by 0.8 to
1.4percent annually.
Almost half the activities people are paid almost
$16trillion in wages to do in the global economy
have the potential to be automated by adapting
currently demonstrated technology, according
to our analysis of more than 2,000 work activities
across 800 occupations. While less than 5percent
of all occupations can be automated entirely using
demonstrated technologies, about 60percent of all
occupations have at least 30percent of constituent
activities that could be automated. More occupations
will change than will be automated away.
Activities most susceptible to automation involve
physical activities in highly structured and predictable
environments, as well as the collection and processing
of data. In the United States, these activities make up
51percent of activities in the economy accounting for
almost $2.7trillion in wages. They are most prevalent
in manufacturing, accommodation and food service,
and retail trade, and include some middle-skill jobs.
Technical, economic, and social factors will determine

the pace and extent of automation. Continued
technical progress, for example in areas such
as natural language processing, is a key factor.
Beyond technical feasibility, the cost of technology,
competition with labor including skills and supply and
demand dynamics, performance benefits including
and beyond labor cost savings, and social and
regulatory acceptance will affect the pace and scope
of automation. Our scenarios suggest that half of
todays work activities could be automated by 2055,
but this could happen up to 20years earlier or later
depending on the various factors, in addition to other
wider economic conditions.
People will need to continue working alongside
machines to produce the growth in per capita GDP
to which countries around the world aspire. Our
productivity estimates assume that people displaced
by automation will find other employment. The
anticipated shift in the activities in the labor force is
of a similar order of magnitude as the long-term shift
away from agriculture and decreases in manufacturing
share of employment in the United States, both of
which were accompanied by the creation of new types
of work not foreseen at the time.
For business, the performance benefits of automation
are relatively clear, but the issues are more
complicated for policy-makers. They should embrace
the opportunity for their economies to benefit from the
productivity growth potential and put in place policies
to encourage investment and market incentives to
encourage continued progress and innovation. At the
same time, they must evolve and innovate policies
that help workers and institutions adapt to the impact
on employment. This will likely include rethinking
education and training, income support and safety
nets, as well as transition support for those dislocated.
Individuals in the workplace will need to engage
more comprehensively with machines as part of their
everyday activities, and acquire new skills that will be
in demand in the new automation age.
A global force that will transform economies and the workforce

Technical automation potential by adapting currently demonstrated technologies
While few occupations are fully automatable, 60 percent of all occupations
have at least 30 percent technically automatable activities
ACTIVITIES WITH HIGHEST
AUTOMATION POTENTIAL:
Predictable physical activities 81%
Processing data
69%
Collecting data
64%
About 60% of occupations

have at least 30% of
their activities that
are automatable
Remaining
countries
Share of roles
100% = 820 roles
2.7
5.1
100% =
$15.8T
100
91
1.1
India
73
1.1
Japan
62
<5% of occupations consist
of activities that are
100% automatable
26
18
Wages associated with technically

automatable activities
$ trillion
United States
100
1.7
Labor associated with technically

Million full-time equivalents (FTEs)
51
42
34
Remaining
countries 332
100% =
>90
>80
>70
>60
>50
>40
>30
China
Big 5 in Europe1
4.1
>20
>10
>0
Japan
Big 5 in Europe1
35
54
China
1.1B
60
233
United States
Technical automation potential, %
394
India
France, Germany, Italy, Spain, and the United Kingdom.
Automation will boost global

productivity and raise GDP
Five factors affecting pace and

extent of adoption
G19 plus Nigeria
TECHNICAL
FEASIBILITY
Technology
has to be
invented,
integrated,
and adapted
into solutions
for specific
case use
COST OF
DEVELOPING
AND
DEPLOYING
SOLUTIONS
Hardware
and software
costs
LABOR
MARKET
DYNAMICS
The supply,
demand, and
costs of
human labor
affect which
activities will
be automated
ECONOMIC
BENEFITS
Include higher
throughput
and increased
quality,
alongside
labor cost
savings
REGULATORY
AND SOCIAL
ACCEPTANCE
Even when
automation
makes
business
sense,
adoption can
take time
Productivity growth, %
Automation can help provide some of the productivity needed
to achieve future economic growth
Employment growth, %
will slow drastically because of aging
Last 50 years Next 50 years

Growth aspiration
Next 50 years
Potential impact of automation
Scenarios around time spent on current work activities, %

Adoption,
Early scenario
Adoption,
Late scenario
Technical automation
potential, Early scenario
100
Technical
automation
potential
must precede
adoption
80
60
40
20
0
potential, Late scenario
2020
2030
2040
2050
2060 2065
Technical,
economic,
and social
factors affect
pace of
adoption
1.8
2.8
1.4
1.7
3.5
Historical
0.1
2.9
Required to
achieve
projected growth
in GDP per capita
0.8
0.1
1.5
0.1
0.9
Early
scenario
Late
scenario
Starship Technologies
viii
McKinsey Global Institute
EXECUTIVE SUMMARY
Automation is not a new phenomenon, and questions about its promise and effects have
long accompanied its advances. More than a half century ago, US President Lyndon B.
Johnson established a national commission to examine the impact of technology on the
economy and employment, declaring that automation did not have to destroy jobs but can
be the ally of our prosperity if we will just look ahead.1 Many of the same questions have
come to the fore again today, as a result of remarkable recent advances in technologies
including robotics, artificial intelligence (AI), and machine learning. Automation now has
the potential to change the daily work activities of everyone, from miners and landscape
gardeners to commercial bankers, fashion designers, weldersand CEOs. But how quickly
will these technologies become a reality in the workplace? And what will their impact be on
employment and on productivity in the global economy?
Over the past two years, we have been conducting a research program on automation
technologies and their potential effects. Some of our key findings include the following.
We are living in a new automation age in which robots and computers can not only
perform a range of routine physical work activities better and more cheaply than
humans, but are also increasingly capable of accomplishing activities that include
cognitive capabilities. These include making tacit judgments, sensing emotion, or even
drivingactivities that used to be considered too difficult to automate successfully.2
The automation of activities can enable productivity growth and other benefits at both
the level of individual process and businesses, as well as at the level of entire economies,
where productivity acceleration is sorely needed, especially as the share of the workingage population declines in many countries. At a microeconomic level, businesses
everywhere will have an opportunity to capture benefits and achieve competitive
advantage from automation technologies, not just from labor cost reductions, but also
from performance benefits such as increased throughput, higher quality, and decreased
downtime. At a macroeconomic level, based on our scenario modeling, we estimate
automation could raise productivity growth on a global basis by as much as 0.8 to
1.4percent annually.
Our approach to analyzing the potential impact of automation is through a focus on
individual activities rather than entire occupations. Given currently demonstrated
technologies, very few occupationsless than 5percentare candidates for full
automation today, meaning that every activity constituting these occupations is
automated. However, almost every occupation has partial automation potential, as a
significant percentage of its activities could be automated. We estimate that about half
of all the activities people are paid to do in the worlds workforce could potentially be
automated by adapting currently demonstrated technologies.
President Johnson signed the bill creating the National Commission on Technology, Automation, and
Economic Progress, on August 19, 1964. The report was published in 1966. Technology and the American
economy: Report of the National Commission on Technology, Automation and Economic Progress, US
Department of Health, Education and Welfare, February 1966. In December 2016, the White House released
a new report on the same subject, Artificial intelligence, automation, and the economy.
2
In this report we focus on the implications of automation technologies rather than on the technologies
themselves. For a more detailed discussion of machine learning and deep learning technologies see the
corresponding chapter in The age of analytics: Competing in a data-driven world, McKinsey Global Institute,
December 2016.
1
The pace and extent of automation, and thus its impact on workers, will vary across
different activities, occupations, and wage and skill levels. Many workers will continue
to work alongside machines as various activities are automated. Activities that are likely
to be automated earlier include predictable physical activities, especially prevalent in
manufacturing and retail trade, as well as collecting and processing data, which are
activities that exist across the entire spectrum of sectors, skills and wages. Some forms
of automation will be skill-biased, tending to raise the productivity of high-skill workers
even as they reduce the demand for lower-skill and routine-intensive occupations, such
as filing clerks or assembly-line workers.3 Other automation has disproportionately
affected middle-skill workers.4 As technology development makes the activities of both
low-skill and high-skill workers more susceptible to automation, these polarization
effects could be reduced.
Automation will have wide-ranging effects, across geographies and sectors. Although
automation is a global phenomenon, four economiesChina, India, Japan, and the
United Statesaccount for just over half of the total wages and almost two-thirds the
number of employees associated with activities that are technically automatable by
adapting currently demonstrated technologies. Within countries, automation potential
will be affected by their sector mix, and the mix of activities within sectors. For example,
industries such as manufacturing and agriculture include predictable physical activities
that have a high technical potential to be automated, but lower wage rates in some
developing countries could constrain adoption.
Automation will not happen overnight, and five key factors will influence the pace
and extent of its adoption. First is technical feasibility, since the technology has to be
invented, integrated and adapted into solutions that automate specific activities. Second
is the cost of developing and deploying solutions, which affects the business case for
adoption. Third are labor market dynamics, including the supply, demand, and costs of
human labor as an alternative to automation. Fourth are economic benefits, which could
include higher throughput and increased quality, as well as labor cost savings. Finally,
regulatory and social acceptance can affect the rate of adoption even when deployment
makes business sense. Taking all of these factors into account, we estimate it will take
decades for automations effect on current work activities to play out fully. While the
effects of automation might be slow at a macro level within entire sectors or economies,
they could be quite fast at a micro level, for an individual worker whose activities are
automated, or a company whose industry is disrupted by competitors using automation.
While much of the current debate about automation has focused on the potential for
mass unemployment, predicated on a surplus of human labor, the worlds economy will
actually need every erg of human labor working, in addition to the robots, to overcome
demographic aging trends in both developed and developing economies. In other
words, a surplus of human labor is much less likely to occur than a deficit of human
labor, unless automation is deployed widely. However, the nature of work will change. As
processes are transformed by the automation of individual activities, people will perform
activities that are complementary to the work that machines do (and vice versa). These
shifts will change the organization of companies, the structure and bases of competition
of industries, and business models.
For a discussion of skill-biased and unskill-biased technical change see David H. Autor, Frank Levy, and
Richard J. Murnane, The skill content of recent technological change: An empirical explanation, Quarterly
Journal of Economics, November 2003, and Daron Acemoglu and David H. Autor, Skills, tasks, and
technologies: Implications for employment and earnings, in Handbook of Labor Economics, volume 4B,
David Card and Orley Ashenfelter, eds., Elsevier, 2011.
4
David H. Autor, Why are there still so many jobs? The history and future of workplace automation, Journal of
Economic Perspectives, volume 29, number 3, 2015.
3
Executive summary
For business, the performance benefits of automation are relatively clear, but the issues
are more complicated for policy makers. They should embrace the opportunity for their
economies to benefit from the productivity growth potential and put in place policies
to encourage investment and market incentives to encourage continued progress and
innovation. At the same time, they must evolve and innovate policies that help workers
and institutions adapt to the impact on employment. This will likely include rethinking
education and training, income support, and safety nets, as well as transition support for
those dislocated. Individuals in the workplace will need to engage more comprehensively
with machines as part of their everyday activities, and acquire new skills that will be in
demand in the new automation age.
The scale of shifts in the labor force over many decades that automation technologies can
unleash is of a similar order of magnitude to the long-term technology-enabled shifts in the
developed countries workforces away from agriculture in the 20th century. Those shifts
did not result in long-term mass unemployment because they were accompanied by the
creation of new types of work not foreseen at the time. We cannot definitively say whether
historical precedent will be upheld this time. But our analysis shows that humans will still
be needed in the workforce: the total productivity gains we estimate will come about only if
people work alongside machines.
GAUGING AUTOMATION POTENTIAL IN THE GLOBAL WORKPLACE TODAY

The Czech writer Karel Capek coined the word robot almost a century ago, in a 1920
play about factory androids that each do the work of two-and-a-half humans at a fraction
of the cost.5 Science fiction has since become business fact. Robots are commonplace
in manufacturing, and algorithms are playing an ever-larger role in companies from UPS
to Amazon.6 With recent developments in robotics, artificial intelligence, and machine
learning, technologies not only do things that we thought only humans could do, but also
can increasingly do them at superhuman levels of performance. Some robots that are far
more flexibleand a fraction of the costof those used in manufacturing environments
today can be trained by frontline staff to perform tasks that were previously thought to be
too difficult for machines, and are even starting to take over service activities, from cooking
hamburgers to dispensing drugs in hospital pharmacies. Artificial intelligence is also making
major strides: in one recent test, computers were able to read lips with 95percent accuracy,
outperforming professional human lip readers who tested at 52percent accuracy.7
We used the state of technology in respect to 18 performance capabilities to estimate
the technical automation potential of more than 2,000 work activities from more than 800
occupations across the US economy, and then broadened our analysis across the global
economy (see BoxE1, How we established the technical automation potential of the
global economy).
The word robot comes from the Slavic word for work, robota. Karel Capek, R.U.R. (Rossums Universal
Robots), 1920. The play is available at www.gutenberg.org.
6
Steven Rosenbush and Laura Stevens, At UPS, the algorithm is the driver, Wall Street Journal, February
16, 2015. Amazon employees can pick and pack three times as many products per hour with the help of
robots. Eugene Kim, Amazon is now using a whole lot more of the robots from the company it bought for
$775million, Business Insider, October 22, 2015; Kim Bhasin and Patrick Clark, How Amazon triggered a
robot arms race, Bloomberg, June 29, 2016.
7
Hal Hodson, Googles DeepMind AI can lip-read TV shows better than a pro, New Scientist, November
21, 2016.
5
A future that works: Automation, employment, and productivity
BoxE1. How we established the technical automation potential of the global economy
To assess the technical automation potential of the global
This analysis enabled us to estimate the technical
economy, we used a disaggregation of occupations into
automation potential of more than 2,000 work activities in
constituent activities that people are paid to do in the
more than 800 occupations across the economy, based
global workplace. Each of these activities requires some
on data from the US Department of Labor. By estimating
combination of 18 performance capabilities, which we list
the amount of time spent on each of these work activities,
in ExhibitE1. They are in five groups: sensory perception,
we were able to estimate the automation potential of
cognitive capabilities, natural language processing, social
occupations in sectors across the economy, comparing
and emotional capabilities, and physical capabilities.
them with hourly wage levels. Drawing on industry
experts, we also developed scenarios for how rapidly the
We estimated the level of performance for each of these
performance of automation technologies could improve in
capabilities that is required to perform each work activity
each of these capabilities.
successfully, based on the way humans currently perform
activitiesthat is, whether the capability is required at
The analysis we conducted for the United States provided
all, and if so, whether the required level of performance
us with a template for estimating the automation potential
was at roughly a median human level, below median
and creating adoption timing scenarios for 45 other
human level, or at a high human level of performance
economies representing about 80percent of the global
(for example, top 25th percentile). We then assessed the
workforce. For details of our methodology, see the
performance of existing technologies today based on the
technical appendix.
same criteria.
Exhibit E1
To assess the technical potential of automation, we structure our analysis around 2,000 distinct work activities
Occupations
Activities
Capability requirements
Sensory perception
Sensory perception
Greet customers
Retail
salespeople
Answer questions
about products
and services
Food and
beverage service
workers
Clean and
maintain
work areas
Teachers
Demonstrate
product features
Health
practitioners
Process sales and

transactions
~800 occupations
~2,000 activities
assessed across all
occupations
SOURCE: McKinsey Global Institute analysis
Executive summary
Cognitive capabilities
Retrieving information
Recognizing known patterns/categories (supervised learning)
Generating novel patterns/categories
Logical reasoning/problem solving
Optimizing and planning
Creativity
Articulating/display output
Coordination with multiple agents
Natural language processing
Natural language generation
Natural language understanding
Social and emotional capabilities
Social and emotional sensing
Social and emotional reasoning
Emotional and social output
Physical capabilities
Fine motor skills/dexterity
Gross motor skills
Navigation
Mobility
Only a small percentage of occupations can be fully automated by adapting

current technologies, but some work activities of almost all occupations could
be automated
Unlike some other studies, the core of our analysis focuses on work activities rather than
whole occupations. We consider work activities a more relevant and useful measure since
occupations are made up of a range of activities with different potential for automation.8 For
example, a retail salesperson will spend some time interacting with customers, stocking
shelves, or ringing up sales. Each of these activities is distinct and requires different
capabilities to perform successfully.
Overall, we estimate that 49percent of the activities that people are paid to do in the
global economy have the potential to be automated by adapting currently demonstrated
technology. While less than 5percent of occupations can be fully automated, about
60percent have at least 30percent of activities that can technically be automated
(ExhibitE2). While certain categories of activity, such as processing or collecting data, or
performing physical activities and operating machinery in a predictable environment, have
a high technical potential for automation, the susceptibility is significantly lower for other
activities including interfacing with stakeholders, applying expertise to decision making,
planning, and creative tasks, or managing and developing people (ExhibitE3).
Exhibit E2
While few occupations are fully automatable, 60 percent of all occupations have at least 30 percent technically
Automation potential based on demonstrated technology of occupation titles in the United States (cumulative)1
Share of roles (%)
100% = 820 roles
Example occupations
100
Sewing machine operators,

graders and sorters of
agricultural products
Technical automation potential (%)
>90
Stock clerks, travel agents,

watch repairers
Chemical technicians,
nursing assistants,
Web developers
Fashion designers, chief

executives, statisticians
>80
<5% of occupations consist of

activities that are 100% automatable
8
18
>70
>60
>50
26
34
42
>40
>30
>20
51

have at least 30% of their
activities that are automatable
62
73
>10
Psychiatrists, legislators
91
>0
100
1 We define automation potential according to the work activities that can be automated by adapting currently demonstrated technology.
SOURCE: US Bureau of Labor Statistics; McKinsey Global Institute analysis
Most recent studies estimating the impact of automation in the workplace focus on occupations. See
Carl Benedikt Frey and Michael A. Osborne, The future of employment: How susceptible are jobs to
computerisation? Oxford Martin School, September 17, 2013; The future of jobs: Employment, skills, and
workforce strategy for the fourth Industrial Revolution, World Economic Forum, January 2016.
REPEATS in report
Exhibit E3
Three categories of work activities have significantly higher technical automation potential
Time spent on activities that can be automated by adapting currently demonstrated technology
%
81
69
64
Time spent
in all US
occupations
%
Total wages
in US, 2014
$ billion
20
14
16
12
17
16
18
Interface3
Unpredictable
physical4
Collect
data
Process
data
Predictable
physical5
896
504
1,030
931
766
Manage1 Expertise2
596
26
18
1,190
Most
susceptible
activities
51%
of total
working hours
$2.7 trillion
in wages
1 Managing and developing people.

2 Applying expertise to decision making, planning, and creative tasks.
3 Interfacing with stakeholders.
4 Performing physical activities and operating machinery in unpredictable environments.
5 Performing physical activities and operating machinery in predictable environments.
NOTE: Numbers may not sum due to rounding.
REPEATS
The degree of automation potential varies considerably among sectors

and countries
A significant degree of variation among sectors of the economy, and among the
occupations within those sectors, emerges from this analysis. For example, almost one-fifth
in
report
of the time spent in US workplaces involves predictable physical activity and is prevalent in
such sectors as manufacturing and retail trade. Accordingly, these sectors have a relatively
high technical potential for automation using todays technology. ExhibitE4 shows a range
of sectors in the US economy broken down into different categories of work activity.9
An interactive mapping of the automation potential of multiple sectors of the economy is available online at
http://public.tableau.com/profile/mckinsey.analytics#!/
Executive summary
Exhibit E4
Technical potential for automation across sectors varies depending on mix of activity types
Size of bubble indicates % of
time spent in US occupations
Ability to automate (%)

0
Sectors by
activity type
Manage Expertise
Interface
Unpredictable
Collect
physical
data
Process
data
Predictable
physical
50
100
Automation potential
%
Accommodation
and food services
73
Manufacturing
60
Agriculture
58
Transportation and
warehousing
57
53
Retail trade
51
Mining
Other services
49
Construction
47
Utilities
44
Wholesale trade
44
Finance and
insurance
43
Arts, entertainment,
and recreation
41
Real estate
40
Administrative
39
Health care and

social assistances
36
Information
36
Professionals
35
Management
35
Educational
services
27
Within sectors, too, there is considerable variation. In manufacturing, for example,

occupations that have a large proportion of physical activities in predictable environments
such as factory welders, cutters, and solderers have a technical automation potential above
90percent based on adapting currently developed technologies, whereas for customer
service representatives that susceptibility is less than 30percent.
While wage and skill levels are negatively correlated with technical automation potential
(on average, occupations with higher wages and skill requirements have lower automation
potential, reflecting some skill bias), a large amount of variation underlies the averages.
Essentially all occupations, whether high skill or low skill, have some technical automation
potential, including CEOs; we estimate about 25percent of their work could potentially
be automated, primarily such tasks as analyzing reports and data to inform decisions,
reviewing status reports, preparing staff assignments, and so on.
At a global level, technically automatable activities touch the equivalent of 1.1billion
employees and $15.8trillion in wages (ExhibitE5). Four economiesChina, India, Japan,
and the United Statesaccount for just over half of these total wages and employees; China
and India together account for the largest technically automatable employment potential
more than 700million full-time equivalents between thembecause of the relative size of
their labor forces. The potential is also large in Europe: according to our analysis, 54million
full-time employee equivalents and more than $1.7trillion in wages are associated with
technically automatable activities in the five largest economiesFrance, Germany, Italy,
Spain, and the United Kingdom.
Our analysis of the technical automation potential of the global economy shows that there is
a range among countries of about 15 percentage points. Two factors explain this range. The
first is the sectoral makeup of each economy, that is, the proportion of a national economy
that is in sectors such as manufacturing or accommodation and food services, both of
which have relatively high automation potential, compared with the proportion in sectors
with lower automation potential such as education. The second factor is the occupational
makeup of sectors in different countries, in other words, the extent to which workers in these
sectors are engaged in job titles with high automation potential, such as manufacturing
production, and those in job titles with lower automation potential such as management and
administration. A detailed look at all 46 countries we have examined is available online.10
The data visualization can be found on the McKinsey Global Institute public site at tableau.com:
10
Executive summary
Exhibit E5
The technical automation potential of the global economy is significant, although there is some variation
among countries
Employee weighted overall % of activities that can be automated
by adapting currently demonstrated technologies1
<45
4547 4749 4951
>51
No data
Technical automation potential is concentrated in countries with the largest populations and/or high wages
Potential impact due to automation, adapting currently demonstrated technology (46 countries)
Wages associated with
technically automatable activities
$ trillion
Labor associated with

Million FTE
%
Japan
United
States
Remaining
countries
55
Remaining
countries
2.7
5.1
100% =
$15.8 trillion
4.1
1.1
India
1.1
Japan
1.7
Europe Big 52
China
India
332
100% =
1,109 million
FTEs
35
Japan 54
60
Europe Big 52
United States
394
China
233
India
52
China
51
United States
46
Europe Big 52
46
Rest of world
50
1 Pakistan, Bangladesh, Vietnam, and Iran are largest countries by population not included.
2 France, Germany, Italy, Spain, and the United Kingdom.
SOURCE: Oxford Economic Forecasts; Emsi database; US Bureau of Labor Statistics; McKinsey Global Institute analysis
FACTORS AFFECTING PACE AND EXTENT OF AUTOMATION

While the technology is advancing, the journey from technical automation potential to full
adoption is nonetheless likely to take decades. The timing is affected by five sets of factors:
Technical feasibility. Technology has to be invented, integrated and adapted into
solutions that automate specific activities. Deployment in the workplace can begin
only when machines have reached the required level of performance in the capabilities
required to carry out particular activities. While machines can already match or
outperform humans on some of the 18 capabilities in our framework, including
information retrieval, gross motor skills, and optimization and planning, many other
capabilities require more technological development. In particular, advancements in
natural language understanding could unlock significantly more technical automation
potential. Emotional and social reasoning capabilities will also need to become more
sophisticated for many work activities. For typical work activities, multiple capabilities,
such as sensory perception and mobility, will be needed simultaneously, and thus
solutions that integrate specific capabilities in context must be engineered.
Cost of developing and deploying solutions. The cost of automation affects the
business case for adoption. Developing and engineering automation technologies
takes capital. Hardware solutions range from standard computers to highly designed,
application-specific hardware such as robots with arms and other moving parts requiring
dexterity. Cameras and sensors are needed for any activity requiring sensory perception
capabilities, while mobility requires wheels or other hardware that enable machines
to move. Such attributes increase costs relative to a general-purpose hardware
platform. Even virtual solutions that are based on software require real investments in
engineering to create solutions. For deployment, hardware requires significant capital
spending, and thus automation that requires it has high initial costs compared to wages.
Software solutions, by comparison, tend to have a minimal marginal cost, which usually
makes them less expensive than wages and thus they tend to be adopted earlier. Over
time, both hardware and software costs decline, making solutions competitive with
human labor for an increasing number of activities.
Labor market dynamics. The quality (for instance, skills), quantity, as well as supply,
demand, and costs of human labor as an alternative affect which activities will be
automated. For example, restaurant cooking has high automation potential, more than
75percent, based on currently demonstrated technologies, but the decision to deploy
the technology will need to take into account the wage costs of cooks, who earn $11 per
hour on average in the United States, and the abundance of people willing to working as
cooks at that wage. Labor market dynamics also differ by geography, not only in terms
of how different and evolving demographics affect the base supply of labor, but also
different wage rates. Manufacturing automation is more likely to be adopted sooner in
countries with high manufacturing wages, such as North America and Western Europe,
than in developing countries with lower wages. Furthermore, the effects of automation
can interact with labor market skills and supply. For example, if middle-income workers
such as clerks and factory workers are displaced by the automation of data collection
and processing and predictable physical activities, they could find themselves moving
into lower paid occupations, increasing supply, and potentially putting downward
pressure on wages. Conversely, they might take time to retrain into other highskill positions, delaying their re-entry into the labor force, and temporarily reducing
labor supply.
Economic benefits. In addition to labor cost savings, a business case for automation
could include performance gains such as increased profit, increased throughput and
productivity, improved safety, and higher quality, which sometimes exceed the benefits
10
Executive summary
of labor substitution (see BoxE2, Automation technologies could provide significant

performance benefits for companies beyond labor substitution). For example, the
benefits of increased production and lower overall maintenance costs by automating the
control room of an oil and gas production facility dwarf those associated with reduced
labor costs in the control room. Automated driving of cars and trucks could not only
reduce the labor costs associated with drivers; it could also potentially improve safety
(the vast majority of accidents are the result of driver errors) and fuel efficiency.
Box E2. Automation technologies could provide significant performance

benefits for companies beyond labor substitution
The deployment of automation technologies could bring a range of performance benefits for
companies. These benefits are varied, depending on the individual use case, and potentially
very substantialin some cases, considerably larger than cost reductions associated with
labor substitution. They include, but are not limited to, greater throughput, higher quality,
improved safety, reduced variability, a reduction of waste, and higher customer satisfaction.
We developed several hypothetical case studies to gain a better understanding of the
potential for automation in different settings and sought to quantify the economic impact
of realizing this vision. The case studies are of a hospital emergency department, aircraft
maintenance, oil and gas operations, a grocery store, and mortgage brokering. The
resultswhile forward-lookingare nonetheless striking. The value of the potential benefits
of automation, calculated as a percentage of operating costs, ranges from between 10-15
percent for a hospital emergency department and a grocery store, to 25 percent for aircraft
maintenance, and more than 90 percent for mortgage origination.
We also see automation being deployed today that is already generating real value. For
example, Rio Tinto has deployed automated haul trucks and drilling machines at its mines
in Pilbara, Australia, and says it is seeing 1020 percent increases in utilization there as
a result.1 Google has applied artificial intelligence from its DeepMind machine learning to
its own data centers, cutting the amount of energy they use by 40 percent.2 In financial
services, automation in the form of straight-through processing, where transaction
workflows are digitized end-to-end, can increase the scalability of transaction throughput
by 80 percent, while concurrently reducing errors by half.3 Safety is another area that could
benefit from increased automation. For example, of the approximately 35,000 road death in
the United States annually, about 94 percent are the result of human error or choice.4
The relative cost of automation can be modest compared with the value it can create. The
types and sizes of investment needed to automate will differ by industry and sector. For
example, industries with high capital intensity that require substantial hardware solutions to
automate and are subject to heavy safety regulation will likely see longer lags between the
time of investment and the benefits than sectors where automation will be mostly softwarebased and less capital-intensive. For the former, this will mean a longer journey to breakeven
on automation investment. However, our analysis suggests that the business case can be
compelling regardless of the degree of capital intensity.
Productivity improvements in a changing world, presentation by Michael Gollschewski, managing director
Pilbara Mines, Rio Tinto, July 13, 2015.
2
Rich Evans and Jim Gao, DeepMind AI reduces energy used for cooling Google data centers by 40%, Google,
blog post, July 20, 2016.
3
Joo Bueno, Yran Bartolumeu Dias, Alexandre Sawaya, Jorge Valadas. End-to-end digitization for securities
services, McKinsey on Payments, September 2014.
4
Barack Obama: Self-driving, yes, but also safe, Pittsburgh Post-Gazette, September 19, 2016.
1
11
Regulatory and social acceptance. Even when deploying automation makes business
sense, the rate of adoption can be affected by contextual factors such as regulatory
approval and the reaction of users. There are multiple reasons that technology adoption
does not happen overnight. The shift of capital investment into these new technologies
takes time (in aggregate), as does changing organizational processes and practices
to adapt to new technologies. Reconfiguring supply chains and ecosystems can be
laborious, and regulations sometimes need to change. Government policy can slow
adoption, and different businesses adopt technologies at different rates. Changing the
activities that workers do also requires dedicated effort, even if they are not actively
resisting. And especially in the case of automation, individuals may feel uncomfortable
about a new world where machines replace human interaction in some intimate life
settings, such as a hospital, or in places where machines are expected to make life-anddeath decisions, such as when driving.
Automation adoption will take decades, across a wide range of

possible scenarios
To analyze a range of potential scenarios for the pace at which automation will affect
activities across the global economy, we constructed a model that simplifies the effects of
these five factors into four timing stages: capability development, solution development,
economic feasibility, and final adoption. The S-curve in ExhibitE6 indicates the potential
time range that emerges from our scenario analyses, with the dark blue line representing an
earliest adoption scenario and the light blue line a latest adoption scenario, aggregating
across all of the activities that account for about 80percent of the worlds workforce. For
example, we estimate that adapting currently demonstrated technology has the technical
potential to automate roughly 50percent of the worlds current work activities. While the
date at which this could happen could be around 2055, assuming all the factors are in
place for successful adoption by then, we modeled possible scenarios where that level of
adoption occurs up to almost 20years earlier or later.
Among the first sectors likely to feel the impact of automation will be those that involve
types of activities we categorize as having the highest automation potential today based
on currently demonstrated technology. From a geographical perspective, advanced
economies are also likely to deploy automation ahead of many emerging economies, largely
because of higher wage levels, which make a stronger business case for deployment.
This magnitude of shifts in work activities over multiple decades is not unprecedented. In
the United States, for example, the share of farm employment fell from 40percent in 1900
to 2percent in 2000, while the share of manufacturing employment fell from approximately
25percent in 1950 to less than 10percent in 2010 (ExhibitE7).11 In both cases, new activities
and jobs were created that offset those that disappeared, although it was not possible to
predict what those new activities and jobs would be while these shifts were occurring.
11
12
Stanley Lebergott, Labor force and employment 18001960, in Output, employment, and productivity in
the United States after 1800, Dorothy S. Brady, ed., NBER, 1966; World Bank data; Mack Ott, The growing
share of services in the US economydegeneration or evolution? Federal Reserve Bank of St. Louis Review,
June/July 1987.
Executive summary
Exhibit E6
Automation will be a global force, but adoption will take decades and there is significant uncertainty on timing
Time spent on current work activities1
%
Adoption
Early scenario
Late scenario
potential
Early scenario
Late scenario
100
90
80
70
60
50
40
30
20
10
0
2016 20
25
30
35
40
45
50
60
55
65
70
75
80
85
90
95
2100
1 Forty-six countries used in this calculation, representing about 80% of global labor force.
Exhibit E7
REPEATS in report
Employment in agriculture has fallen from 40 percent in 1900 to less than 2 percent today
Distribution of labor share by sector in the United States, 18402010
%
90
Rest of the economy
80
70
60
50
40
30
20
Manufacturing
10
0
1840 50
Agriculture
60
70
80
90 1900 10
20
30
40
50
60
70
80
90 2000 2010
SOURCE: Stanley Lebergott, Labor force and employment 18001960, in Output, employment, and productivity in the United States after 1800, Dorothy S.
Brady, ed., NBER, 1966; World Data Bank, World Bank Group; FRED: Economic Research, Federal Reserve Bank of St. Louis; Mack Ott, The
growing share of services in the US economydegeneration or evolution? Federal Reserve Bank of St. Louis Review, June/July 1987; McKinsey
Global Institute analysis
REPEATS in report
13
Modeling scenarios for the pace and extent of automation adoption

Capability development is the first stage that we modeled for the timing of automation
adoption. Deployment in the workplace can begin only when machines have reached the
required level of performance in the capabilities required to carry out particular activities.
Once the technical capabilities have been developed, they must be integrated into solutions
that can execute specific activities in context, that is, to create commercially available
systems. Our analysis suggests that, on average, this solution development process can
take between one and nine years.
The third stage we modeled for scenario timelines is when automation is economically
feasible. For modeling purposes, we assume that adoption begins when the developed
solution for any given activity is at or below the cost for human workers to perform that
activity in a specific occupation and within a particular country.12 While the performance
benefits of automation sometimes exceed those related to labor cost savings, our
conservative modeling assumes that decision-makers discount the benefits of initial labor
cost savings by roughly the same amount as they believe the also uncertain non-labor costrelated benefits will be captured.
Adoption and deployment of automation, the fourth stage we modeled to develop
scenarios, can also be a slow process. For our analysis, we looked at the historical adoption
rates a wide range of 25 technologies, involving both hardware and software, as well as
business and consumer technologies. The time between the commercial availability of
these technologies and their eventual maximum level of adoption generally took at least
nearly a decade and in some cases multiple decades, with the time range between eight
and 28years.13
EVEN AS IT CAUSES SHIFTS IN EMPLOYMENT, AUTOMATION CAN GIVE A

STRONG BOOST TO PRODUCTIVITY AND GLOBAL GDP GROWTH
Automation will cause significant labor displacement and could exacerbate a growing skills
and employment gap that already exists between high-skill and low-skill workers.14 Our
analysis of automation potential also suggests that many occupations could be partially
automated before they are fully automated, which could have different implications for highand low-skill workers.15 Especially for low-skill workers, this process could depress wages
unless demand grows. Viewed through a long-term perspective, however, as we described
previously, large-scale historical structural shifts in the workplace where technology has
caused job losses have, over time, been accompanied by the creation of a multitude of
new jobs, activities, and types of work.16 Furthermore, labor markets can be quite dynamic:
almost five million people leave their jobs every month in the United States, of whom about
By costs in this case, we mean wages plus benefits, calculated globally on a purchasing power parity basis.
Some of the technologies we modeled have not likely yet reached their eventual peak in adoption.
14
There has been a proliferation of books by competing schools of techno-optimists and techno-pessimists.
They notably include Erik Brynjolfsson and Andrew McAfee, The second machine age: Work, progress, and
prosperity in a time of brilliant technologies, W. W. Norton & Company, 2014; Robert Gordon, The rise and
fall of American growth: The US standard of living since the Civil War, Princeton University Press, 2016; and
Martin Ford, Rise of the robots: Technology and the threat of a jobless future, Basic Books, 2015. Also Jason
Furman, Is this time different? The opportunities and challenges of artificial intelligence, remarks at AI Now:
The Social and Economic Implications of Artificial Intelligence Technologies in the Near Term conference in
New York, July 7, 2016.
15
For a discussion of the polarization of the labor market, see David H. Autor and David Dorn, The growth of
low-skill service jobs and the polarization of the US labor market, American Economic Review, volume 103,
number 5, August 2013.
16
For example, a study conducted by McKinsey & Companys French office in 2011 showed that for every job
that had been lost in France as the result of the advent of the internet in the previous 15years, 2.4 new jobs
had been created. Impact dinternet sur lconomie franaise: Comment internet transforme notre pays
(The internets impact on the French economy: How the internet is transforming our country), McKinsey &
Company, March 2011.
12
13
14
Executive summary
three million do so voluntarily. Most of these people are not unemployed for long periods as
they move on to other jobs.17
That said, automation also represents a very substantial opportunity to support global
economic growth. Our estimates suggest it has the potential to contribute meaningfully to
the growth necessary to meet the per capita GDP aspirations of every country, at a time
when changing demographics call those aspirations into question. Indeed, for this growth to
take place, rather than having a massive labor surplus, everyone needs to keep working
with the robots working alongside them.
Automation can help close a GDP growth gap resulting from declining growth
rates of working-age populations
GDP growth was brisk over the past half century, driven by the twin engines of employment
growth and rising productivity, both contributing approximately the same amount. However,
declining birthrates and the trend toward aging in many advanced and some emerging
economies mean that peak employment will occur in most countries within 50years.18 The
expected decline in the share of the working-age population will open an economic growth
gap: roughly half of the sources of economic growth from the past half century (employment
growth) will evaporate as populations age. Even at historical rates of productivity growth,
economic growth could be nearly halved.
Automation could compensate for at least some of these demographic trends. We estimate
the productivity injection it could give to the global economy as being between 0.8 and
1.4percent of global GDP annually, assuming that human labor replaced by automation
would rejoin the workforce and be as productive as it was in 2014. Considering the labor
substitution effect alone, we calculate that, by 2065, automation could potentially add
productivity growth in the largest economies in the world (G19 plus Nigeria) that is the
equivalent of an additional 1.1billion to 2.3billion full-time workers (ExhibitE8).
The productivity growth enabled by automation can ensure continued prosperity in aging
nations and provide an additional boost to fast-growing ones. Automation on its own will not
be sufficient to achieve long-term economic growth aspirations across the world; for that,
additional productivity-boosting measures will be needed, including reworking business
processes or developing new products and services.
Potential impact of automation in three groups of countries

Automation could boost productivity and help close the economic growth gap in the 20
largest economies in the medium term, to 2030. We have divided these countries into three
groups, each of which could use automation to further national economic growth objectives,
depending on their demographic trends and growth aspirations. The three groups are:
Advanced economies, including Australia, Canada, France, Germany, Italy, Japan, South
Korea, the United Kingdom, and the United States. These economies typically face an
aging workforce, with the decline in working-age population growth more immediate in
some (Germany, Italy, and Japan) than in others. Automation can provide the productivity
boost required to meet economic growth projections that they otherwise would struggle
to attain without other significant productivity growth accelerators. These economies
thus have a major interest in pursuing rapid automation adoption.
Emerging economies with aging populations. This category includes Argentina, Brazil,
China, and Russia, which face economic growth gaps as a result of projected declines in
US Bureau of Labor Statistics job openings and labor turnover survey database.
Global growth: Can productivity save the day in an aging world? McKinsey Global Institute, January 2015.
Our estimate of employment growths contribution to GDP growth in this report differs slightly from this earlier
research, as we have assumed productivity measured in each country, rather than a global average.
17
18
15
the growth of their working population.19 For these economies, automation can provide
the productivity injection needed just to maintain current GDP per capita. To achieve a
faster growth trajectory that is more commensurate with their developmental aspirations,
these countries would need to supplement automation with additional sources of
productivity, such as process transformations, and would benefit from more rapid
adoption of automation.
Emerging economies with younger populations. These include India, Indonesia, Mexico,
Nigeria, Saudi Arabia, South Africa, and Turkey.20 The continued growth of the workingage population in these countries could support maintaining current GDP per capita.
However, given their high growth aspirations, automation plus additional productivityraising measures will be necessary to sustain their economic development.
Exhibit E8
Globally, automation could become a significant economic growth engine as employment growth wanes
GDP growth for G19 and Nigeria
Compound annual growth rate
%
Historical
19642014
Future
201565
3.5
2.9
Productivity
growth
1.8
2.8
Employment
growth
1.7
1.4
0.2
0.1
GDP per capita growth

%
1.5
0.9
0.8
0.1
0.1
0.1
0.1
Historical
Required to
maintain current
GDP per capita
Required to
Early scenario
Late scenario
achieve projected
GDP per capita
2.1
2.5
0.1
6.7
Full-time equivalent gap1

Billion
0.51.1
1 Additional full-time equivalents (FTEs) needed to achieve growth target.

SOURCE: The Conference Board Total Economy database; United Nations Population Division; McKinsey Global Institute analysis
The demographic trends are pronounced for China and Russia, while Argentinas future workforce gap is
less certain.
20
The populations of Saudi Arabia and Turkey are projected to grow strongly over the next 20years, but
slow thereafter.
19
REPEATS in report
16
Executive summary
The advances in automation and their potential impact on national economies could upend
some prevailing models of development and challenge ideas about globalization. Countries
experiencing population declines or stagnation will be able to maintain living standards
even as the labor force wanes. Meanwhile, countries with high birthrates and a significant
growth in the working-age population may have to worry more about generating new jobs
in a new age of automation. Moreover, low-cost labor may lose some of its edge as an
essential developmental tool for emerging economies, as automation drives down the cost
of manufacturing globally.
HOW BUSINESS LEADERS, POLICY MAKERS, AND WORKERS CAN PREPARE

FOR THE NEW AUTOMATION AGE
Business leaders, policy makers, and workers everywhere face considerable challenges in
capturing the full potential of automations beneficial effect on the economy, even as they
navigate the major uncertainties about the social and employment repercussions.
Automation will give business leaders opportunities to improve their
performance and enter new markets, but they will need to rework their
processes and organizations
Automation of various activities can improve the performance of almost any business
process.21 Beyond enabling reduction in labor costs, automation can raise throughput,
increase reliability, and improve quality, among other performance gains.
To assess where automation could be most profitably applied to improve performance,
business leaders may want to conduct a thorough inventory of their organizations activities
and create a heat map of where automation potential is high. Once they have identified
business processes with activities that have high automation potential, these could be
reimagined to take full advantage of automation technologies (rather than just mechanically
attempting to automate individual activities in the current processes). They could then
assess the benefits and feasibility of these automation-enabled process transformations.
Taking advantage of these transformations could lead to significant displacements in labor.
Business leaders would be well served to consider how to best redeploy that labor, whether
within their own organizations or elsewhere, both to improve their own performance and
to act as good corporate citizens. Retraining and skill-raising programs will be important
to support workers shifting to new roles and taking on new activities. It will also be critical
for corporate leaders to ensure that the organizational elements of their companies are
adapting to the advent of automation.
On a strategic level, automation could enable the emergence of massively scaled
organizations, instantly able to propagate changes that come from headquarters.
Technology will make measuring and monitoring easier, providing effective new tools for
managers. However, greater scale means that errors could be more consequential, which in
turn will require stronger quality controls.
Even as some corporations could be scaling up, automation and digital technologies more
generally will enable small players, including individuals and small companies, to undertake
project work that is now largely carried out within bigger firms. The growth of very small
and very large companies could create a barbell-shaped economy, in which mid-sized
companies lose out. In all sectors, automation could heighten competition, enabling firms
to enter new areas outside their previous core businesses, and creating a growing divide
between technological leaders and laggards in every sector.
We explore several case studies of the potential transformations of business processes in Chapter 3.
21
17
For policy makers, an embrace of automation could be accompanied by

measures to raise skills and promote job creation, and by rethinking incomes
and social safety nets
Policy makers globally will have a strong incentive to encourage and enable rapid adoption
of automation technologies in order to capture the full productivity boost necessary to
support economic growth targets. At the same time, they will need to think through how to
support the redeployment of potentially large numbers of displaced workers, since the full
economic benefits of automation depend on workers continuing to work.22
Early adoption of automation could benefit from policy support, both in regard to
the technology development, and for its deployment. That will require investment in
developing the technologies themselves, and also in digitally enabled infrastructure to
support automation.
Labor redeployment will be one of the most important societal challenges. Governments
are often not particularly adept at anticipating the types of jobs that could be created, or
new industries that will develop. However, they could initiate and foster dialogues about
what work needs doing, and about the grand societal challenges that require more
attention and effort.23 Governments could also seek to encourage new forms of technologyenabled entrepreneurship, and intervene to help workers develop skills best suited for
the automation age. For example, many economies are already facing a shortage of
data scientists and business translators.24 Governments working with the private sector
could take steps to ensure that such gaps are filled, establishing new education and
training possibilities.
One of the challenges of the new era will be to ensure that wages are high enough for the
new types of employment that will be created, to prevent continuing erosion of the wage
share of GDP, which has dropped sharply since the 1970s.25 If automation does result in
greater pressure on many workers wages, some ideas such as earned income tax credits,
universal basic income, conditional transfers, shorter workweeks, and adapted social safety
nets could be considered and tested. As work evolves at higher rates of change among
sectors, locations, activities, and skill requirements, many workers may need assistance in
adjusting to the new age.
Workers will need to work more closely with technology, freeing up more time
to focus on intrinsically human capabilities that machines cannot yet match
Men and women in the workplace will need to engage more comprehensively with
machines as part of their everyday activities. Tighter integration with technology will free up
time for human workers including managers to focus more fully on activities to which they
bring skills that machines have yet to master. This could make work more complex, and
harder to organize, with managers spending more time on coaching.26
As people make education and career choices, it will be important for them to be made
aware of the factors driving automation in particular sectors, to help them identify the skills
Jason Furman, Is this time different? The opportunities and challenges of artificial intelligence, remarks at AI
Now: The Social and Economic Implications of Artificial Intelligence Technologies in the Near Term conference
in New York, July 7, 2016.
23
Tim OReilly, Dont replace people. Augment them, Medium.com, July 17, 2016.
24
The age of analytics: Competing in a data-driven world, McKinsey Global Institute in collaboration with
McKinsey Analytics, December 2016.
25
Poorer than their parents? Flat or falling incomes in advanced economies, McKinsey Global Institute,
July 2016.
26
Daron Acemoglu and Pascual Restrepo, The race between machines and humans: Implications for growth,
factor shares and jobs, Vox, July 5, 2016.
22
18
Executive summary
that could be useful for them to acquire from a labor-market perspective, and what activities
will be complements of activities that are likely to be automated.27
High-skill workers who work closely with technology will likely be in strong demand, and
may be able to take advantage of new opportunities for independent work as the corporate
landscape shifts and project work is outsourced by companies. Middle-skill workers whose
activities have the highest technical potential for automation (predictable physical activities,
collecting and analyzing data) can seek opportunities for retraining to prepare for shifts
in their activities toward those that are complements of activities the machines will start
to perform.
Low-skill workers working with technology will be able to achieve more in terms of output
and productivity but may experience wage pressure given the potentially large supply of
similarly low-skill workers.
Education systems will need to evolve for a changed workplace, with policy makers working
with education providers to improve basic skills in the STEM fields of science, technology,
engineering, and mathematics, and put a new emphasis on creativity, as well as on critical
and systems thinking. For all, developing agility, resilience, and flexibility will be important at
a time when everybodys job is likely to change to some degree.
Finally, automation will create an opportunity for those in work to make use of the innate
human skills that machines have the hardest time replicating: logical thinking and problem
solving, social and emotional capabilities, providing expertise, coaching and developing
others, and creativity. For now, the world of work still expects men and women to undertake
rote tasks that do not stretch these innate capabilities as far as they could. As machines take
on ever more of the predictable activities of the workday, these skills will be at a premium.
Automation could make us all more human.
Automation will play an essential role in providing at least some of the productivity boost that
the global economy needs over the next half century as growth in working-age populations
declines. It will contribute meaningfully to GDP per capita growth, even if it will not on its
own enable emerging economies to meet their fast-growth aspirations. Given the range
of scenarios around the pace and extent of adoption of automation technologies, there
are sure to be surprises. We will see large-scale shifts in workplace activities over the next
century. These trends are already under way. Policy makers, business leaders, and workers
themselves must not wait to take action: already today, there are measures that can be
taken to prepare, so that the global economy can capture the opportunities offered by
automation, even as it avoids the drawbacks.
Erik Brynjolfsson and Andrew McAfee, The second machine age: Work, progress, and prosperity in a time of
brilliant technologies, W. W. Norton & Company, 2014.
27
19
20
Executive summary
OVERVIEW OF SELECT RECENT STUDIES ON THE

IMPACT OF AUTOMATION AND FUTUREOFWORK
Exhibit
E9
Carl Benedikt
Frey
and Michael A.
Osborne
Citibank with
Frey and Osborne
OECD
World Economic
Forum
McKinsey Global
Institute
January 2016
June 2016
January 2016
January 2017
Jobs/occupations
Tasks
Not applicable
Work activities
50+ countries and

regions
21 OECD countries
15 major developed
and emerging
economies
46 countries
representing about 80%
of global labor force
Extension of FreyOsborne (2013), using

World Bank data, to
estimate impact of
automation globally.
Further analyses
include examination of
demographic
changes, global value
chain, etc.
Estimates of
automatibility of tasks
were developed
based on matching of
the automatibility
indicators by FreyOsborne and the
PIAAC data
occupational codes,
followed by a twostep, tailored
regression analysis
Analysis of large-scale
survey of major global
employers, including
100 largest global
employers in each of
WEF main industry
sectors, to estimate
the expected level of
changes in job
families between
201520 and
extrapolate number of
jobs gained/lost
Disaggregation of
occupations into 2,000
constituent activities and
rating each against
human performance in
18 capabilities. Further
analysis of time spent
on each activity and
hourly wage levels.
Scenarios for
development and
adoption of automation
technologies
Date
September 2013
Unit of analysis
Jobs/occupations
Scope
US labor market
Approach summary
Analysis of 702
occupations (70 handlabeled working with
ML researchers,
followed by a tailored
Gaussian process
classifier to estimate
others and confirm
hand-labels) to
approximate the
impact of future
computerization on
the US labor market
Key relevant findings
Building on Frey
About 47% of total
and Osborne's
US occupations are
original work, data
at high risk of
from the World
automation perhaps
Bank suggests the
over the next
risks are higher in
decade or two
Wages and
many other
educational
countries; in the
attainment show a
OECD, on average
strong negative
57% of jobs are
relationship with
susceptible to
probability of
automation. This
computerization
number rises to
69% in India and
77% in China
On average, 9% of Automation and

technological
jobs across the 21
advancements
OECD countries are
could lead to a net
automatable
There are notable
employment impact
differences across
of more than
OECD countries
5.1 million jobs lost
when it comes to
to disruptive labor
automation (e.g.,
market changes
the share of
between 201520,
automatable jobs is
with a total loss of
6% in Korea vs.
7.1 million jobs
12% in Austria)
two-thirds of which
are concentrated in
the office and
administrative job
familyand a total
gain of 2 million
jobs in several
smaller job families
Almost half of work

activities globally have
the potential to be
automated using
current technology.
<5% of occupations
can be automated
entirely; about 60%
have at least 30% of
Technically
touch 1.1 billion
workers and
$15.8 trillion in wages.
China, India, Japan,
and the United States
constitute over half
Automations boost to
global productivity
could be 0.81.4%
annually over
decades
SOURCE: Carl Benedikt Frey and Michael A. Osborne, The future of employment: How susceptible are jobs to computerisation?, Oxford Martin School,
September 17, 2013; Technology at Work v2.0: The future is not what it used to be, Citibank, January 2016; The future of jobs: Employment, skills,
and workforce strategy for the fourth Industrial Revolution, World Economic Forum, January 2016; Melanie Arntz, Terry Gregory, and Ulrich Zierahn,
The risk of automation for jobs in OECD countries: A comparative analysis, OECD Social, Employment and Migration working paper number 189,
OECD, May 2016; McKinsey Global Institute analysis. See also Holger Bonin, Terry Gregory, and Ulrich Zierahn, Ubertragung der Studie von
Frey/Osborne: (2013) auf Deutschland, Forschungsbericht 455, Bundesministerium fr Arbeit und Soziales, April 14, 2015; and Jeremy Bowles, The
computerisation of European jobs, Bruegel, July 24, 2014.
21
Computers can read lips more accurately than experienced humans.

Burton Pritzker/The Image Bank/Getty Images
22
Executive summary
1. THE NEW FRONTIER

It is easy to become blas about technological progress in this non-stop, 24/7, digitaleverything-always-and-everywhere era. We take technological advances almost for granted
and are frustrated when an app that streams the latest Hollywood movies crashes, or a
smartphone which has many times the processing power of a 1980s Cray 2 supercomputer
does not fire up the moment we press the on button.
We forget that it was not always this easy. Not so long ago, we had to go to libraries to look
up quotations and insert compact discs into an audio system to play music. Transmitting
even tiny amounts of data was complicated by todays standards; sometimes we had to
strap bulky acoustic couplers onto a fixed telephone and wait for the modem to screech. As
the German sociologist Hartmut Rosa has pointed out, we live in an age of acceleration in
which the art of saving time has reached unprecedented heights thanks to technology, but
we nonetheless feel that we must run faster just to stay put.28
Even by these standards, however, some of the most recent developments in robotics,
artificial intelligence, and machine learning are noteworthy for the advances they represent.
We are on the cusp of a new automation age in which technologies not only do things that
we thought only humans could do, but also can increasingly do them at a superhuman level
of performance. In this report, we focus on the adoption and implications of automation
technologies rather than on the technologies themselves. However, by way of introduction,
this chapter lays out some key areas of recent technical advancesand where remaining
technical obstacles must still be overcome to achieve the full promise of workplace
automation.29
AUTOMATION TECHNOLOGIES ARE INCREASINGLY

OUTPERFORMING HUMANS
Physical robots have been around for a long time in manufacturing, but now we are seeing
much more flexible, safer, and less expensive robots engaging in service activitiesand
improving over time as they are trained by their human coworkers on the shop floor.30 For
example, some hospitals now regularly use automated systems for storing and dispensing
medication in their pharmacies, eliminating human picking errors, and also have automated
haul and transport for their clinical supplies.31 The advances in cognitive tasks are no
less striking. Software has long been able to outperform humans in some areas, such as
financial-service transactions or route optimization for companies such as UPS.32 Now,
artificial intelligence is starting to encroach on activities that were previously assumed to
require human judgment and experience. Exhibit1 is a non-exhaustive list of some of the
technologies and techniques that are being developed to enable automation of different
work activities.
Hartmut Rosa, Social acceleration: A new theory of modernity, Columbia University Press, 2013.
For a more detailed discussion of machine learning and deep learning technologies see the corresponding
chapter in The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016.
30
Baxter robots by Rethink Robotics can now pick up items that are not precisely aligned, and then reorient and
place them correctly. http://www.rethinkrobotics.com/baxter/what-makes-our-robots-different.
31
Swisslog offers a medication management system, www.swisslog.com/en/products. For a transport and
logistics example, see www.aethon.com/tug/tughealthcare.
32
For example, insurance companies use novel pattern recognition to detect fraudulent claims, saving
companies including GE millions of dollars annually. A 2014 study used US Securities and Exchange
Commission filing data as well as social network analysis to determine clusters of insiders and correlated their
trading patterns. See Tamersov Acar et al., Large-scale insider-trading analysis: Patterns and discoveries,
Georgia Institute of Technology, August 2014.
28
29
Exhibit 1
Glossary of automation technologies and techniques
JANUARY 2017
This list is not comprehensive but is meant to illustrate some of the technologies and techniques
that are being developed to enable automation of different work activities
Technologies
and techniques
Artificial
intelligence
Description/examples
Field of computer science specializing in developing systems that exhibit intelligence. Often
abbreviated as AI, the term was coined by John McCarthy at the Dartmouth Conference in 1956, the
first conference devoted to this topic
Machine
learning
Subfield of artificial intelligence developing systems that learn, i.e., practitioners

train these systems rather than programming them
Supervised
learning
Machine learning techniques that train a system to respond appropriately to stimuli by

providing a training set of sample input and desired output pairs. Supervised learning
has been used for email spam detection by training systems on a large number of
emails, each of which has been manually labeled as either being spam or not
Transfer
learning
Subfield of machine learning developing systems that store knowledge gained while
solving one problem and applying it to a different but related problem. Often used
when the training set for one problem is small, but the training data for a related
problem is plentiful, e.g., repurposing a deep learning system trained on a large nonmedical image data set to recognize tumors in radiology scans
Reinforcement
learning
Subfield of machine learning developing systems that are trained by receiving virtual
rewards or punishments for behaviors rather than supervised learning on correct
input-output pairs. In February 2015, DeepMind described a reinforcement learning
system that learned how to play a variety of Atari computer games. In March 2016,
DeepMinds AlphaGo system defeated the world champion in the game of Go
Cognitive
computing
Synonym for artificial intelligence
Neural networks Artificial

neural
network
AI systems based on simulating connected neural units, loosely modeling the way
that neurons interact in the brain. Computational models inspired by neural
connections have been studied since the 1940s
Deep
learning
Use of neural networks that have many layers (deep) of a large number (millions) of
artificial neurons. Prior to deep learning, artificial neural networks often only had three
layers and dozens of neurons; deep learning networks often have seven to ten or more
layers. The term was first used in 2000
Convolutional neural
network
Artificial neural networks in which the connections between neural layers are inspired
by the organization of the animal visual cortex, the portion of the brain that processes
images, well suited for perceptual tasks. In 2012, the only entry using a convolutional
neural network achieved an 84% correct score in the ImageNet visual recognition
contest, vs. a winning score of 75% the year prior. Since then, convolutional neural
networks have won all subsequent ImageNet contests, exceeding human performance
in 2015, above 90%
Recurrent
neural
network
Artificial neural networks whose connections between neurons include loops, wellsuited for processing sequences of inputs. In November 2016, Oxford University
researchers reported that a system based on recurrent neural networks (and
convolutional neural networks) had achieved 95% accuracy in reading lips,
outperforming experienced human lip readers, who tested at 52% accuracy.
SOURCE: John McCarthy et al, A proposal for the Dartmouth summer research project on artificial intelligence, August 31, 1955, AI Magazine, volume 27,
number 4, 2016; Hayit Greenspan, Bram van Ginneken, and Ronald M. Summers, Deep learning in medical imaging: Overview and future promise
of an exciting new technique, IEEE Transactions on Medical Imaging, volume 35, number 5, May 2016; Volodymyr Mnih, Human-level control
through deep reinforcement learning, Nature, February 25, 2015; Igor Aizenberg, Naum N. Aizenberg, and Joos P.L. Vandewalle, Multi-valued and
universal binary neurons: Theory, learning and applications, Springer Science & Business Media, 2000; www.image-net.org; Yannis M. Assael et al,
LipNet: End-to-end sentence-level lipreading, University of Oxford (forthcoming); McKinsey Global Institute analysis
24
1. The new frontier
35
Exhibit 1
Glossary of automation technologies and techniques (continued)
JANUARY 2017
This list is not comprehensive but is meant to illustrate some of the technologies and techniques
that are being developed to enable automation of different work activities
Technologies
and techniques
Description/examples
Robotics
Soft robotics
Non-rigid robots constructed with soft and deformable materials that can manipulate
items of varying size, shape and weight with a single device. Soft Robotics Inc.
grippers can adaptively pick up soft foods (e.g., baked goods, tomatoes) without
damaging them.
Swarm
robotics
Coordinated multi-robot systems, often involving large numbers of mostly

physical robots
Tactile/touch
robotics
Robotic body parts (often biologically inspired hands) with capability to sense, touch,
exhibit dexterity, and perform variety of tasks
Serpentine
robots
Serpentine looking robots with many internal degrees of freedom to thread through
tightly packed spaces
Humanoid
robots
Robots physical similar to human beings (often bi-pedal) that integrate variety of AI
and robotics technologies and are capable of performing variety of human tasks
(including movement across terrains, object recognition, speech, emotion sensing,
etc.). Aldebaran Robotics and Softbanks humanoid Pepper robot is being used to
provide customer service in more than 140 Softbank Mobile stores in Japan
Autonomous
cars and
trucks
Wheeled vehicles capable of operating without a human driver. In July 2016, Tesla
reported that its cars had driven over 130 million miles while on Autopilot. In
December 2016, Rio Tinto had a fleet of 73 driverless trucks hauling iron ore 24
hours/day in mines in Western Australia
Unmanned
aerial
vehicles
Flying vehicles capable of operating without a human pilot. The unarmed General
Atomics Predator XP UAV, with roughly half the wingspan of a Boeing 737, can fly
autonomously for up to 35 hours from take-off to landing
Chatbots
AI systems designed to simulate conversation with human users, particularly those

integrated into messaging apps. In December 2015, the General Services
Administration of the US Government described how it uses a chatbot named Mrs.
Landingham (a character from the television show The West Wing) to help onboard
new employees
Robotic
process
automation
Class of software robots that replicates the actions of a human being interacting with
the user interfaces of other software systems. Enables the automation of many backoffice (e.g., finance, human resources) workflows without requiring expensive IT
integration. For example, many workflows simply require data to be transferred from
one system to another
Automation
product
categories
SOURCE: www.ald.softbankrobotics.com; A tragic loss, Tesla blog, June 30, 2016; Resource revolution: Transformations beyond the supercycle, McKinsey
Global Institute, forthcoming in 2017; www.ga-asi.com/predator-xp; Jessie Young, How a bot named Dolores Landingham transformed 18Fs
onboarding, www.18f.gsa.gov, December 15, 2015; McKinsey Global Institute analysis
25
Some of AIs exploits are less heralded than its victory over a human champion of the
complex board game Go in March 2016.33 For example, a project by Googles DeepMind
and the University of Oxford has applied deep learning to a huge data set of BBC programs
to create a lip-reading system. Trained using more than 5,000 hours of BBC TV programs,
containing more than 100,000 sentences, it easily outperformed a professional human
lipreader. In tests on 200 randomly selected clips, the professional annotated just
12.4percent of words without error, while the computer annotated 46.8percent errorfreeand many of its mistakes were small ones, such as leaving a plural s off the end of a
word.34
Many other surprising technologies are making advances. Robot skin made of a
piezoelectronic transistor mesh developed by Georgia Tech and covered in thousands of
mechanical hairs is as sensitive as human skin and able to feel textures and find objects
by touch.35 In the social and emotional realm, Affectiva, a Boston-based company, uses
advanced facial analysis to monitor emotional responses to advertisements and other digital
media content, via a webcam.36 In the United Kingdom, the University of Hertfordshire
has developed a minimally expressive humanoid robot called KASPAR that operates as a
therapeutic toy for children with autism. Having physical, human-like properties, yet being
non-human, allows the children to investigate the human-looking featuresfor example,
squeezing KASPARs nose or tickling its toessafely and in a way that would not be
possible or appropriate with a real person.37
Such advances suggest that an idea toyed with by science fiction writers for at least a
centurythat of robots and other machines replacing men and women in the workplace on
a large scalecould soon become a reality. We seem to be approaching a new frontier, but
we have not arrived there quite yet.
WHERE THE BIGGEST REMAINING OBSTACLES LIE

The Czech writer Karel Capek first used the word robot in 1920, in a play about a factory in
which androids created partly through a chemical process each do the work of two-and-ahalf humans at a fraction of the cost. One of his characters explains: Robots are not people.
Mechanically they are more perfect than we are, they have an enormously developed
intelligence, but they have no soul.38 Today, mechanical perfection seems achievable, as
robots become ever more adept at physical tasks, even if they are still wobbly on uneven
terrain and consume a lot of energy. Through deep reinforcement learning, they can also
untie shoelaces, unscrew bottle caps, and remove a nail from the back of a hammer.39
Their intelligence, too, has progressedbut this is where the most formidable technical
challenges still lie ahead. While machines can be trained to perform a range of cognitive
tasks, they remain limited. They are not yet good at putting knowledge into context, let alone
improvising. They have little of the common sense that is the essence of human experience
and emotion. They struggle to operate without a pre-defined methodology. They are far
more literal than people, and poor at picking up social or emotional cues. Sarcasm and irony
Choe Sang-Hun, Googles computer program beats Lee Se-dol in Go tournament, New York Times, March
15, 2016.
34
Hal Hodson, Googles DeepMind AI can lip-read TV shows better than a pro, New Scientist, November
21, 2016.
35
Klint Finley, Syntouch is giving robots the ability to feel textures like humans do, Wired, December 17, 2015.
36
http://www.affectiva.com/solutions/overview/.
37
Ricky Boleto, Could robots help children with autism? BBC News, March 10, 2014. See also the University
of Hertfordshire web page for KASPAR, http://www.herts.ac.uk/kaspar/supporting-children-with-autism.
38
The word robot comes from robota, the Slavic word for work. Karel Capek, R.U.R. (Rossums Universal
Robots), 1920. The play is available at www.gutenberg.org. Capek initially called the creatures labori but was
persuaded by his brother to change the name. Science diction: The origin of the word robot, NPR Science
Friday, April 22, 2011.
39
Signe Brewster, A strong robot hand with a softer side, MIT Technology Review, February 9, 2016; Robots
master skills with deep learning technique, Kurzweil, May 22, 2015.
33
26
1. The new frontier
pass them by. They generally cannot detect whether a customer is upset at a hospital bill or
a death in the family, and for now, they cannot answer What do you think about the people
in this photograph? or other open-ended questions. They can tell jokes without really
understanding them. They dont yet feel humiliation, fear, pride, anger, or happiness. They
also struggle with disambiguation, unsure whether a mention of the word mercury refers to
a planet, a metal, or the winged god of Roman mythology.
Moreover, while machines can replicate individual performance capabilities such as
fine motor skills or navigation, much work remains to be done integrating these different
capabilities into holistic solutions where everything works together seamlessly. Combining
a range of technologies will be essential for workplace automation, but engineering such
solutionswhether for hardware or softwareis a difficult process. The creation of
solutions that solve specific problems in the workplace is work that will have to be done as
individual technical challenges are overcome in the lab. Even once the technical feasibility
issues have been resolved and the technologies become commercially available, it can take
years before they are adopted.
Yet, given the speed with which technological advances are happening, reaching and
crossing the next frontier may just be a question of time. Moores lawthat the number of
transistors in a dense integrated circuit doubles approximately every two yearsmay be
slowing, but we are still seeing massive increases in computing power. Machine learning
and its subset deep learning continue to advance rapidly, while traditional AI algorithms
become more versatile and powerful. Cloud computing and other technologies are opening
new possibilities for more people to become involved in innovation. Academic research
in these areas, especially in artificial intelligence, has increased significantly, and global
markets are taking notice, with growing corporate investment in research and development.
When large-scale automation does come to the workplace, what will that mean for the
economy, for jobs, and for the future of work itself? And how fast could it happen?
Such existential questions are easier to answer through fiction. Capeks 1920 play about
robots ends with the destruction of mankind and robots discovering the meaning of love.
This report, by contrast, seeks to establish a fact base with which to address these issues
and a foundation for a more informed dialogue. Robots may not have a soul, but their
potential impact on the global economy can be calculated.
27
A man works on a robot at a digital factory in The Hague that produces carbon plates for the aircraft industry.
Bart Maat/European Pressphoto Agency b.v./Alamy Stock Photo
28
1. The new frontier
2. THE TECHNICAL POTENTIAL

FORAUTOMATION
The speed with which automation technologies are emerging, and the extent to which
they could disrupt the world of work, may appear daunting but is not unprecedented.
Technological change has reshaped the workplace continually over the past two centuries
since the Industrial Revolution, and even earlier (see Box1, What history teaches us about
the effect of technological change on work, employment, and productivity). Nonetheless,
the latest technological developments will touch every job in every sector and in every
country. The advances we described in the previous chapter all have practical applications
in the workplace, and in some cases they have already been adapted, integrated, and
deployed. Sophisticated machines are replacing human labor in workplaces from factories
to fast-food restaurants. They are becoming a part of everyday life in fields from journalism
to law to medicine; at the University of Tokyo, for example, IBMs Watson made headlines
in 2016 by diagnosing in a 60-year-old woman a rare form of leukemia that had eluded her
doctors for months.40
These technologies bring with them progress, productivity improvements, increased
efficiencies, safety, and convenience, but they also raise difficult questions about the
broader impact of automation on the workforce as a whole. Think tanks and organizations
such as the World Economic Forum are forecasting the likelihood of major job substitution
by automation.41 Some academic studies estimate that close to 50percent of US and
European jobs could be automated, although other studies put that figure much lower (see
the table Overview of select recent studies on the impact of automation and the future of
work, page21).
Some of these projections focus on occupations perceived to be at risk.42 Our approach
is substantially different: we consider work activities a more relevant and useful basis for
analysis than occupations. The reason for this is that, within sectors, every occupation
consists of a number of constituent activities that may have a different technical potential for
automation. A typical retail salesperson, for example, will spend some time interacting with
customers, stocking shelves, or ringing up sales. Machines can already outperform humans
in some of these activitiesthey are highly adept at managing warehouse inventory, for
exampleand at least one fashion company has a bot that advises clients, via their mobile
phones, about the best lipstick match.43 But computers and machines are far less adept
than humans at sensing the emotional state of customers or understanding context. For
example, no robot yet has the capacity to sense a distressed client and propose offering him
or her a glass of water or a cup of tea.
Bernie Monegain, IBM Watson pinpoints rare form of leukemia after doctors misdiagnosed patient,
Healthcare IT News, August 8, 2016.
41
The World Economic Forum has predicted that more than five million jobs could be lost to robots in 15 major
developed and emerging economies over the next five years. The future of jobs: Employment, skills, and
workforce strategy for the fourth Industrial Revolution, World Economic Forum, January 2016.
42
The Oxford University study, for example, focuses on occupations that it categorizes as being susceptible to
automation. Carl Benedikt Frey and Michael A. Osborne, The future of employment: How susceptible are jobs
to computerization? Oxford Martin School, September 17, 2013.
43
Sephora debuts two new bot-powered beauty tools for Messenger, PR Newswire, November 2, 2016.
40
Box1. What history teaches us about the effect of technological

change on work, employment, and productivity
The fear of technological innovation destroying jobs and displacing workers
dates back several hundred years, even before the Luddite movement in Britain
during the Industrial Revolution that gave its name to militant technophobia. The
Luddites were textile mill workers in Nottingham who rioted in 1811 to destroy
the new automated looms that threatened their livelihoods. Ever since, there has
been no shortage of predictions that machines would replace human laborers,
with possibly dire effects. Karl Marx wrote in 1858 that the means of labor passes
through different metamorphoses, whose culmination is the machine, or rather,
an automatic system of machinery.1
In 1930, the British economist John Maynard Keynes coined the term
technological unemployment to describe a situation in which innovation that
economized the use of labor outstripped the pace at which new jobs could
be created. Keynes warned that this was akin to a new diseasebut he also
described this malady as being a temporary phase of maladjustment.2
More recently, in 1966, a report from the US National Commission on Technology,
Automation, and Economic Progress, predicted that in the new technology,
machines and automated processes will do the routine and mechanical work.
Human resources will be released and available for new activities beyond those
that are required for mere subsistence. The great need is to discover the nature of
this new kind of work, to plan it, and to do it. In the longer run, significant changes
may be needed in our societyin education, for exampleto help people find
constructive and rewarding ways to use increasing leisure.3
One lesson of history is that deployment of new technologies in the past has led to
new forms of work, including in cases when shifts in the activities performed in the
workplace have been very substantial. In the United States, for example, the share
of farm employment fell from 40percent in 1900 to 2percent in 2000; similarly,
the share of manufacturing employment fell from 25percent in 1950 to less than
10percent in 2010.4 In both cases, while some jobs disappeared, new ones
were created, although what those new jobs would be could not be predicted at
the time.
Technological innovation can create new demand and whole new industries.
Printing is one example. When the Times of London in 1814 switched to a
revolutionary steam-powered printing press invented by German engineer
Friedrich Koenig, the newspapers printers staged a revolt that was quelled only
when the paper promised to keep on displaced workers. That prototype, which
used steam from water heated by coal to drive the press, initially printed 1,100
pages per hour, or five times as many as the mechanical press that preceded it.
By 1820, presses could print 2,000 sheets per hour. By 1828, that doubled to
4,000. Then came the invention of rotary presses, which in turn enabled huge rolls
of paper to be loaded into the presses rather than individual sheets. By the 1860s,
Karl Marx, Grundrisse: Foundations of the critique of political economy, 1858 (unpublished
manuscript), available online at www.marxists.org/archive.
2
John Maynard Keynes, Economic possibilities for our grandchildren, in Essays in Persuasion,
Macmillan, 1933. The essay is available online at www.econ.yale.edu/smith/econ116a/keynes1.
pdf.
3
Technology and the American economy: Report of the National Commission on Technology,
Automation, and Economic Progress, US Department of Health, Education, and Welfare, 1966.
4
Stanley Lebergott, Labor force and employment 18001960, in Output, employment,
and productivity in the United States after 1800, Dorothy S. Brady, ed., NBER, 1966; World
Bank data.
1
30
2. The technical potential forautomation
Box1. What history teaches us about the effect of technological

change on work, employment, and productivity (continued)
the most advanced presses could print 30,000 pages per hour. The arrival of
electricity and the development of linotype and photomechanical processes able
to reproduce photographs meant that by 1890, the New York Herald was able to
print 90,000 copies of its four-page paper per hour, with color illustrations. This
stream of innovation, combined with greater press freedom, drove the growth of
a vibrant and fast-growing newspaper industry in the United States and Europe,
creating millions of jobs in printing, journalism, and other related fields.5
More recent evidence at a macroeconomic level suggests the positive links
between technological progress, productivity, and jobs continued through the
20th century. Positive gains in both productivity and employment have been
reported in the United States in more than two-thirds of the years since 1929.6
One-third of new jobs created in the United States in the past 25years did not
exist, or barely existed, 25years ago.7 However, in recent years, there has been a
notable divergence between productivity and pay, and the labor share of income
has declined in many advanced economies.8
The question today is whether this latest wave of innovation is by its nature
substantially different from technological disruptions in the past. As automation
makes inroads into the workplace, a critical concern is that technology-enabled
automation could replace not just low-skill jobswhich is what happened in the
pastbut that it could affect all jobs. For now, there is an increasing bifurcation
in the labor market between a dwindling number of high-skill jobs and many
low-wage and low-skill service jobs.9 As we detail later in this chapter, even
high-paying occupations in sectors such as financial services are potentially
susceptible to automation. Opinions are sharply divided about the medium- and
long-term effects of this automation wave. In 2014, the Pew Research Center
conducted a survey of technology professionals and economists and found that
48percent of respondents believed new technologies would displace more jobs
than they would create by 2025.10
Elizabeth Eisenstein, The printing press as an agent of change, Cambridge University Press,
1980; Robert Hoe, A short history of the printing press and of the improvements in printing
machinery from the time of Gutenberg up to the present day, 1902.
6
Growth and renewal in the United States: Retooling Americas economic engine, McKinsey
Global Institute, February 2011.
7
Ibid.
8
Josh Bivens and Lawrence Mishal, Understanding the historic divergence between productivity
and a typical workers pay: Why it matters and why its real, Economic Policy Institute briefing
paper number 406, September 2015; Poorer than their parents? Flat or falling incomes in
advanced economies, McKinsey Global Institute, July 2016.
9
See David H. Autor and David Dorn, The growth of low-skill service jobs and the polarization
of the US labor market, American Economic Review, volume 103, number 5, August 2013,
and Lawrence Mishel and Kar-Fai Gee, Why arent workers benefiting from labour productivity
growth in the United States? International Productivity Monitor, number 23, spring 2012.
10
Aaron Smith and Janna Anderson, AI, robotics, and the future of jobs, Pew Research Center,
August 6, 2014. There has been a proliferation of books by competing schools of technooptimists and techno-pessimists. They notably include Erik Brynjolfsson and Andrew McAfee,
The second machine age: Work, progress, and prosperity in a time of brilliant technologies, W.
W. Norton & Company, 2014; and Robert Gordon, The rise and fall of American growth: The
US standard of living since the Civil War, Princeton University Press, 2016; Martin Ford, Rise of
the robots: Technology and the threat of a jobless future, Basic Books, 2015. Also see Jason
Furman, Is this time different? The opportunities and challenges of artificial intelligence, remarks
at AI Now: The Social and Economic Implications of Artificial Intelligence Technologies in the Near
Term conference in New York, July 7, 2016.
5
31
Using data from the US Bureau of Labor Statistics and O*Net, we have examined in detail
more than 2,000 work activities for more than 800 occupations across the entire economy.
We estimated the amount of time spent on these activities and the technical feasibility
of automating each of them by adapting currently demonstrated technology. Having
undertaken this analysis of the US economy, we extended our study to 45 other countries,
using the most comparable data available in each.44 This detailed research enables us
to draw important conclusions about the technical feasibility of automation for the global
economy today, as well as for individual professions within specific sectors, from US
mortgage brokers to Indian farmers.
Our core findings are that the proportion of occupations that can be fully automated by
adapting currently demonstrated technologyin other words, all of their activities could
be automatedis very small, less than 5percent in the United States. Automation will
nonetheless affect almost all occupations, not just factory workers and clerks, but also
landscape gardeners and dental lab technicians, fashion designers, insurance sales
representatives, and CEOs, to a greater or lesser degree. The automation potential of
these occupations depends on the types of work activity that they entail, but as a rule of
thumb, about 60percent of all occupations have at least 30percent of activities that are
technically automatable.
In the United States, the country for which we have the most complete data, about
46percent of time spent on work activities across occupations and industries is technically
automatable based on currently demonstrated technologies. Exhibit2 shows the
distribution range of this automation potential in the United States. On a global scale, we
calculate that the adaptation of currently demonstrated automation technologies could
affect 49percent of working hours in the global economy. This potential corresponds to the
equivalent of 1.1billion workers and $11.9trillion in wages. Among countries, the potential
ranges between 40 and 55percent, with just four countriesChina, India, Japan, and the
United Statesaccounting for just over half the total wages and workers. The potential
could also be large in Europe: according to our analysis, the equivalent of 54million full-time
workers and more than $1.9trillion in wages are associated with technically automatable
activities in the continents five largest economies aloneFrance, Germany, Italy, Spain, and
the United Kingdom.
In this chapter we describe in detail the technical potential for automation in different sectors
of the economy and for the global economy as a whole, based on the state of technology
today. We explain how we calculate technical automation potential and, based on that
methodology, we identify categories of activities that are the most and the least susceptible
to automation. This enables us to provide detailed estimates of the technical automation
potential of sectors and of different occupations within those sectors. We conclude with an
examination of similarities and differences between countries, both advanced and emerging
economies, for a global view of automation and its very substantial potential to transform the
world of work.
44
32
For full details of our methodology, see the technical appendix.
Exhibit 2
While few occupations are fully automatable, 60 percent of all occupations have at least 30 percent technically
Automation potential based on demonstrated technology of occupation titles in the United States (cumulative)1
Share of roles (%)
100% = 820 roles
Example occupations
100
Sewing machine operators,

graders and sorters of
agricultural products
Chemical technicians,
nursing assistants,
Web developers
Fashion designers, chief

executives, statisticians
Technical automation potential (%)
Stock clerks, travel agents,

watch repairers
>90
>80
<5% of occupations consist of

activities that are 100% automatable
8
18
>70
>60
>50
26
34
42
>40
>30
>20
51

have at least 30% of their
activities that are automatable
62
73
>10
Psychiatrists, legislators
>0
91
100
1 We define automation potential according to the work activities that can be automated by adapting currently demonstrated technology.
ESTIMATING THE TECHNICAL POTENTIAL FOR WORKPLACE AUTOMATION

Humans at work carry out a wide variety of activitieswhether stapling documents,
examining spreadsheets, meeting clients, interviewing potential new recruits, lifting crates in
a store, or planting corn in a fieldwithout consciously analyzing the exact skill sets they are
DUPLICATE
from
ESof these actions requires a combination of innate or acquired capabilities,
using. In
fact, each
ranging from manual dexterity to social perceptiveness. In order to understand and map
performance requirements for machines in the workplace, we developed a detailed
framework of 18 human capabilities, including physical, emotional, and cognitive ones.
It is important to note that when we discuss automation potential in this chapter, we refer
to the technical potential for automation by adapting technologies that have already been
demonstrated. As the technology becomes more advanced, that potential will also evolve.
While this chapter focuses on technologies that have been developed today, later chapters
estimate the speed with which the technological capabilities are likely to improve and be
adopted in the workplace.
33
It should also be noted that technical feasibility is not the same as actual implementation
of automation, nor a complete predictor that an activity will be automated. A second factor
to consider is the cost of developing and deploying both the hardware and the software
solutions for automation. The cost of labor and related supply and demand dynamics
represent a third factor: if workers are in abundant supply and significantly less expensive
than automation, this could be a decisive argument against adoption. A fourth factor to
consider is the benefits beyond labor substitution, including higher levels of output, better
quality, and fewer errors. The economic value of these benefits is often larger than that
of reducing labor costs. Regulatory and social acceptance issues, such as the degree to
which machines are acceptable in any particular setting, must also be weighed. A robot
may, in theory, be able to replace some of the functions of a nurse, for example. But for now,
the prospect that this might actually happen in a highly visible way could prove unpalatable
for many patients, who expect and trust human contact. The pace at which automation will
take hold in a sector or occupation reflects a subtle interplay between these factors and the
trade-offs among them.
A framework of capabilities to understand the performance requirements of

work activities
The framework of 18 capabilities that we developed to assess automation potential
addresses a wide range of performance requirements. Many of these capabilities
correspond to the technologies we discussed in the previous chapter. They cover five
areas: sensory perception, cognitive capabilities, natural language processing, social and
emotional capabilities, and physical capabilities (Exhibit3).
Sensory perception. This includes visual perception, tactile sensing, and auditory
sensing, and involves complex external perception through integrating and analyzing
data from various sensors in the physical world.
Cognitive capabilities. A range of capabilities is included in this category including
recognizing known patterns and categories (other than through sensory perception);
creating and recognizing novel patterns and categories; logical reasoning and problem
solving using contextual information and increasingly complex input variables;
optimization and planning to achieve specific objectives given various constraints;
creating diverse and novel ideas or a novel combination of ideas; information retrieval,
which involves searching and retrieving information from a large range of sources;
coordination with multiple agents, which involves interacting with other machines and
with humans to coordinate group activity; and output articulation and presentation,
which involves delivering outputs other than through natural language. These could be
automated production of pictures, diagrams, graphs, or mixed media presentations.
Natural language processing. This consists of two distinct parts: natural language
generation, which is the ability to deliver spoken messages, including with nuanced
human interaction and gestures, and natural language understanding, which is
the comprehension of language and nuanced linguistic communication in all its
rich complexity.
Social and emotional capabilities. This consists of three types of capability: social
and emotional sensing, which involves identifying a persons social and emotional state;
social and emotional reasoning, which entails accurately drawing conclusions based
on a persons social and emotional state, and determining an appropriate response;
and social and emotional output, which is the production of an appropriate social or
emotional response, both in words and through body language.
34
Physical capabilities. This includes gross motor skills, fine motor skills, navigation,
and mobility. These capabilities could be implemented by robots or other machines
manipulating objects with dexterity and sensitivity, moving objects with multidimensional
motor skills, autonomously navigating in various environments, and moving within and
across various environments and terrain.
Exhibit 3
Current technologies have achieved different levels of human performance across 18 capabilities
Below median
Automation capability
Median
Top quartile
Capability
level1
Description (ability to )
Sensory
perception
Sensory perception
Autonomously infer and integrate complex external perception

using sensors
Cognitive
capabilities
Recognizing known
patterns/categories
(supervised learning)
Recognize simple/complex known patterns and categories

other than sensory perception
Generating novel patterns/

categories
Create and recognize new patterns/categories (e.g.,

hypothesized categories)
Logical reasoning/ problem

solving
Solve problems in an organized way using contextual

information and increasingly complex input variables other
than optimization and planning
Optimization and planning
Optimize and plan for objective outcomes across various

constraints
Creativity
Create diverse and novel ideas, or novel combinations of ideas
Information retrieval
Search and retrieve information from a large scale of sources

(breadth, depth, and degree of integration)
Coordination with multiple

agents
Interact with others, including humans, to coordinate group

activity
Output articulation/
presentation
Deliver outputs/visualizations across a variety of mediums

other than natural language
Natural
language
processing
Deliver messages in natural language, including nuanced

human interaction and some quasi language (e.g., gestures)
Natural language
understanding
Comprehend language, including nuanced human interaction
Social and
emotional
capabilities
Identify social and emotional state
Social and emotional

reasoning
Accurately draw conclusions about social and emotional state,

and determine appropriate response/action
Social and emotional output
Produce emotionally appropriate output (e.g., speech, body

language)
Manipulate objects with dexterity and sensitivity
Gross motor skills
Move objects with multidimensional motor skills
Navigation
Autonomously navigate in various environments
Mobility
Move within and across various environments and terrain
Physical
capabilities
1 Assumes technical capabilities demonstrated in commercial products, R&D, and academic settings; compared against human performance.
35
We estimated the level of performance in each of these capabilities that is required to

successfully perform each work activity, categorizing whether the capability is required at
all, and if so, whether the required level of performance is at roughly a median human level,
below median human level, or at a high level of performance (for example, the top 25th
percentile).45 We also assessed the performance of existing technologies today against the
same criteria.
This framework enabled us to assess the state of technology today and the potential to
automate work activities in all sectors of the economy by adapting currently demonstrated
technologies. By evaluating the technologies across a spectrum of performance, we have
also been able to take into account their potential evolution in the future and resulting
incremental effect on workplace activities. A detailed account of our methodology is
contained in the technical appendix. Our assessments are simplifications for modeling
purposes that synthesize a variety of subcapabilities, not all of which consistently fall into the
below-median, median, and top-quartile categories.
Machines will need to be able to use many of these capabilities together in the
workplace, as humans do
The 18 capabilities we have identified should not be taken in isolation. They are
closely interconnected.
Let us return to our retail salesperson, by way of example, to show the interplay of these
capabilities. Daily activities may include greeting customers, answering questions about
products and services, cleaning and maintaining work areas, demonstrating product
features, and processing sales and transactions. To carry out this range of activities requires
almost the full spectrum of these capabilities. It starts with the greeting of customers. A
skilled salesperson will identify the social and emotional state of a customer, accurately
draw conclusions about how to react to that social and emotional state, and through body
language, tone of voice, and choice of vocabulary, provide an emotionally appropriate
response. Cognitive capabilities will be fully used, too. Listening to what a customer says
and responding requires the ability to understand and generate natural language. Other
cognitive capabilities employed are the ability to retrieve information (do we have these
shoes in stock?); to reason logically and solve problems (we dont have them in your size
in black, but we do have them in red or brown); to coordinate with multiple agents (Ill
have one of my colleagues determine if we have the item in stock), and creativity (try the
purple pair, theyre very fashionable this year and will suit you well). Physical capabilities are
likewise also needed. They include mobility and navigation (walking to the stockroom), gross
motor skills (taking the shoe box off a shelf in the stockroom), and fine motor skills (tying
a lace).
For now, automation technologies do not have this full range of capabilities to perform at the
same level as humans, and for many problems, they have yet to be seamlessly integrated
into solutions. Combinations of social, cognitive, and physical capabilities are required for
many activities, and we do see patterns about which capabilities are often required together
(Exhibit4). For example, an activity that needs any social or emotional capability typically
needs all of themsensing, reasoning, and output. Likewise, an activity requiring one form
of physical capability such as fine motor skills or mobility tends to require multiple physical
capabilities, including gross motor skills and navigation. However, cognitive capabilities can
occur in combination with many different capabilities. There is an overlap between creativity
and optimization and planning, for example, but one can do without the other. Coordinating
with multiple agents does not automatically entail information retrieval, logical reasoning, or
45
36
Among the many factors that define levels of performance are the acceptable error rates, particularly for
sensory and cognitive activities. For instance, the consequences of false positives or negatives when making
certain law enforcement or health-care judgments could be much more significant than for other judgments in
the entertainment industry.
generating novel patterns. Moreover, where there are cognitive demands, physical demands
are less likely to be required.
Exhibit 4
Many capabilities tend to be required together for specific activities
Social
and
emotional
Reasoning
81 100
Sensing
80
89 100
42
44
46 100
41
42
43
69 100
Information
retrieval
22
24
24
51
42 100
Recognizing
known patterns
14
18
18
42
45
42 100
Generating novel
patterns
14
14
27
17
33
31 100
19
26
24
41
37
42
48
49 100
15
18
17
28
27
12
28
19
32 100
Creativity
16
20
22
23
20
23
26
34
35
20 100
Output articulation/display
10
10
35
37
42
31
15
24
10
24 100
Coordination with
33
multiple agents
31
31
12
15
11
-12
28
-1
Perception
16
13
13
-29 -26 -19 -14 -11 -24 -11 -12 -20
24 100
Fine motor
skills/dexterity
-3
-5
-6
-40 -39 -28 -25 -13 -21 -12 -15 -31
10
44 100
Gross motor
skills
-10 -13 -13 -50 -48 -37 -33 -18 -25 -16 -15 -36
11
43
79 100
Mobility
Navigation
Fine motor skills/

dexterity
Gross motor
skills
Perception
Coordination with
multiple agents
Creativity
Optimization and
planning
Logical reasoning
Information
retrieval
Recognizing
known patterns
Generating novel
patterns
Generation
Understanding
100
Logical
Cogni- reasoning
tive
Optimization and
planning
Physical
Sensory Physical
Cognitive
Output
Natural Understanding
language Generation
Sensory
Sensing
High
Output
Low
Natural
language
Social and
emotional
Reasoning
Correlation between
capability scores
among activities
%
An activity needing any social or

emotional capability; typically
needs a higher score in all of them
-9
Cognitive activities require many

different mental capabilities
-5
Physical requirements are

less likely to be required when
there are cognitive demands
100
Navigation
-13 -17 -12
-4
-8
-17
12
27
36
32 100
Mobility
-20 -18 -18 -11 -13 -13
-4
-9
-17
36
44
49
60 100
When machines can take on workplace activities, the nature of work will change. Today,
only about 10percent or less of the average human workers time at work is spent using
capabilities such as emotional reasoning and creativity, which many people would describe
as being a core part of the human experience. The capability most used is recognizing
known patterns, followed by natural language generation (for example, speaking), sensory
37
perception, information retrieval, and natural language understanding (Exhibit5). By allowing

machines to handle more mundane activities, automation could free up women and men to
use their creative and other talents more than they do now.
Exhibit 5
Recognizing known patterns and natural language generation are the two most-used capabilities in work activities
Time spent by US workers on activities that require median or higher levels of human performance
for each capability
% of time
Recognizing known patterns
67
46
41
Sensory perception
38
35

Gross motor skills
17
15
13
13
12
11
10
Emotional and social output
10
9

Navigation
Mobility
Creativity
Gauging the automation potential of occupations and sectors of the economy

through analysis of their constituent activities
Applying our methodology to the more than 2,000 activities across all sectors of
the economy carried out by the US workforceand subsequently adopting similar
methodology for the global workforcewe found that many types of activities share
common characteristics that are readily grouped into categories. For example, processing
data is a very frequent activity common to a large range of occupations in different sectors,
as is carrying out repetitive physical movement. By analyzing the amount of time spent
on each of these categories of activity, we were able to estimate the technical automation
potential of hundreds of occupations across the economy. We have also analyzed the
implications across wage rates. As noted, the technical ability to automate is only one
element that will lead to automation actually being deployed in the workplace; given
38
hardware and software costs and relative wage levels, a coherent business case needs to
be made, and regulatory, social, and organizational issues also play a role.
Hourly wage rates are not strong predictors of automation potential

Occupations across the spectrum of the economy, from CEO to metal welders, have a range
of automation potential based on todays technologies. Technically automatable activities
represent about $2.7trillion of addressable wages in the United States, or about 46percent
of the total hours worked. Automation is sometimes depicted as primarily affecting particular
groups of workers depending on their wage levels. Our analysis finds that while there is a
negative correlation between wage rates and technical automation potential, there is a large
amount of variation, so the hourly wage rate is not a strong predictor of technical automation
potential.46 In fact, a significant proportion of highly paid work, not just low-wage work, can
be automated (Exhibit6).
Exhibit 6
Both low and high-wage occupations have significant technical automation potential
Automatability1
%
100
90
File
clerks
80
70
60
50
40
30
Chief
executives
20
Landscaping
workers
10
0
10
15
25
20
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100 105
Hourly wage
$
1 Our analysis used detailed work activities, as defined by O*NET, a program sponsored by the US Department of Labor, Employment and Training
Administration.
People in the lowest wage group, earning less than $15 per hour, carry out work activities
that have among the highest potential for automation, and as a whole a majority (51percent)
of this groups activities is automatable. However, those earning between $15 and $30
per hour have an automation potential of 46percent, which is close to the average for
the economy as a whole. Above this $15 to $30 wage level, there is no clear pattern or
correlation between wages and the automation potential of the work. More than 17million
Using a linear model, we find the correlation between wages and automatability (the percentage of time spent
on activities that can be automated by adapting currently demonstrated technology) in the US economy to be
significant (p-value < 0.01), but with a high degree of variability (r2 = 0.19).
46
39
American workers earn between $30 and $45 per hour, but the automation potential of this
group is similar to that of people earning $90 to $105 per hour, for example.
Similarly, higher education attainment levels and work experience are correlated with lower
technical automation potential, but there is also a great deal of variability.
Technical potential is an important consideration, but other factors including

cost of automation and wage levels also weigh on the decision to automate
The deployment of automation technologies in the workplace depends on a range of
factors, of which the availability of the technology itself is an important one, but not the only
one. That it is technically possible for a robot to carry out an activity does not mean that it
will necessarily be deployed to do so in the workplace. Four other factors also need to be
taken into consideration. The pace at which activities are automated and the extent of that
automation reflects a subtle interplay between these factors and the trade-offs among them.
The first factor is the cost. Buying, adapting, and integrating the necessary hardware
and software to automate activities can be expensive and complex, and before doing so,
employers need a strong business case. Will machines in fact be able to undertake the
tasks at hand less expensively and more efficiently, or do human labor and skills still have
the edge? While it may be possible to automate service at a fast-food restaurant, the cost
of the machines compared to the cost of humans earning a minimum wage will need to be
calculated and considered. Indeed, a more clear-cut case for automation may come as
higher-wage jobs are reviewed for their technical feasibility to be automated.
The cost of labor and the related supply and demand dynamics may thus play a significant
role in decisions about automation, and are the second factor. If workers are in abundant
supply and significantly less expensive than automation, this could be a decisive argument
against automation. We calculate that just over $1trillion in wages could be economically
automated with a technology cost of $20 per hour, and $2trillion could be captured with an
automation cost of $10 per hour. Our analysis at the level of individual activities supports the
argument that some occupations in the middle of the income and skill distribution are more
susceptible to automation than others at the top and bottom (see Box2, Labor market
polarization and the technical automation potential of occupation families).
Box2. Labor market polarization and the technical automation potential of occupation families
A body of work in the economics literature documents
occupations) for transportation, office administration,
polarization in the labor markets of developed
and production, in the middle of the distribution, have
countries, in which wage gains went disproportionately
higher percentages of activities with high technical
to those at the top and at the bottom of the income
automation potential (collecting data, processing data,
1
These
and skill distribution, not to those in the middle.
and predictable physical activities) than other occupation
observations have largely been based on workforce
families lower and higher in the distribution. However,
data from national statistical agencies. Our analysis at
one occupation family at the low end of the income
the level of individual activities supports the argument
distribution, food preparation, has the highest percentage
that some occupations in the middle of the income and
of time in activities with a high technical potential for
skill distribution are more susceptible to automation than
automation. Furthermore, as technology continues to
others at the top and bottom. As depicted in Exhibit7,
develop over time, the automation potential of different
occupation families (an aggregation of individual
activities will also increase.
David H. Autor, Why are there still so many jobs? The history and future of workplace automation, Journal of Economic Perspectives, volume
29, number 3, 2015.
40
Box2. Labor market polarization and the technical automation potential of occupation families
(continued)
Exhibit 7
Mix of activity types in selected occupation families
Activity categories with

technical automation potential
% of activity hours spent arranged by weighted hourly wage in $

2
4
2
2
14
2
3
12
21
2
6
11
0
6
32
20
High
Automation
potential1
%
25
Low
24
34
Managing people
31
Applying
expertise
18
14
Interfacing with
stakeholders
20
39
0
5
15
83
0
14
23
23
35
58
17
33
Unpredictable
26
physical activities
Collecting data
64
Processing data
69
24
14
10
13
0
Occupation
family2
58
12
5
Personal
care
Building/ Transgrounds porcleaning tation
Office
administration
Produc- Sales
tion
Management
Weighted
10.41
hourly wage
$
11.76
12.45
16.12
16.56
16.98
17.79
49.24
FTEs
employed
Million
3.89
4.26
8.85
20.96
8.39
13.95
6.54
Food
preparation
11.85
Predictable
81
physical activities
1 Data for the United States only.

2 Aggregations of individual occupations. Eight occupation families selected from a total of 22.
Box 2
41
A third factor to consider are the benefits of automation beyond labor substitution. These
can include higher levels of output, raised quality, speed, agility, safety, and fewer errors.
The potential savings from these benefits can be larger than those from labor costs.
Finally, there is the issue of social and regulatory acceptability of automation. While selfdriving autos and trucks are undergoing tests in both the United States and Europe, they
will be able to operate without human co-drivers only when regulators are comfortable with
them doing so. Social acceptance may be even more difficult. While a robot in theory could
carry out some functions of a nurse or a home-care help, the human beings on the receiving
end of their care may balk at the idea.
Differentiating work activities into seven high-level categories

Across the more than 2,000 work activities across the US economy that we analyzed using
Bureau of Labor Statistics data, we identified seven high-level categories of work activity.
Each of these categories has a different potential for automation. Three categories have
the highest technical potential for automation: performing physical activity and operating
machinery in predictable environments, processing data, and collecting data. The other
four high-level categories have a considerably lower potential for automation: performing
physical activities and operating machinery in unpredictable environments; interfacing with
stakeholders; applying expertise to decision making, planning, and creative tasks; and, least
susceptible to automation, managing and developing people (Exhibit8).
Exhibit 8
Three categories of work activities have significantly higher technical automation potential
Time spent on activities that can be automated by adapting currently demonstrated technology
%
81
69
64
Time spent
in all US
occupations
%
Total wages
in US, 2014
$ billion
20
14
16
12
17
16
18
Interface3
Unpredictable
physical4
Collect
data
Process
data
Predictable
physical5
896
504
1,030
931
766
Manage1 Expertise2
596
26
18
1,190
Most
susceptible
activities
1 Managing and developing people.
2 Applying expertise to decision making, planning, and creative tasks.
3 Interfacing with stakeholders.
4 Performing physical activities and operating machinery in unpredictable environments.
5 Performing physical activities and operating machinery in predictable environments.
42
51%
of total
working hours
$2.7 trillion
in wages
Performing physical activities and operating machinery in

predictable environments
Almost one-fifth of the time spent in US workplaces involves performing physical activities
and operating machinery in predictable environments, which have the highest automation
potential of our seven categories, 81percent. Currently demonstrated technology works
best in these environments, where changes are relatively easy to anticipate. Physical
activities in predictable environments figure prominently in sectors such as manufacturing,
accommodation and food services, and retailing. That makes these sectors among the
most susceptible to automation. Exhibit9 shows a breakdown of different sectors of the
economy based on the seven high-level categories.
In manufacturing, for example, performing physical activities or operating machinery in a
predictable environment represents one-third of the workers overall time. The activities
range from packaging products to loading materials on production equipment to welding
to maintaining equipment. Because of the prevalence of such predictable physical work,
almost 60percent of all manufacturing activities could be automated. The overall potential,
however, masks considerable variance. Within manufacturing, welders, cutters, solderers,
and brazers, have an automation potential above 90percent, for example, while that of
customer service representatives is below 30percent.
Manufacturing is the second most readily automatable sector in the US economy. A service
sector occupies the top spot: accommodation and food services, where almost half of
all labor time involves physical activities in predictable environments and the operation of
machineryincluding preparing, cooking, or serving food; cleaning food preparation areas;
and preparing hot and cold beverages. According to our analysis, 73percent of the activities
workers perform in the accommodation and food services sector have the technical
potential for automation.
Some of this potential is familiar. Automats, or automated cafeterias, for example, have
long been in use. Now restaurants are testing new, more sophisticated concepts, such
as self-service ordering or even robotic servers. Solutions such as Momentum Machines
hamburger-cooking robot, which can reportedly assemble and cook 400 burgers an hour,
could automate a number of cooking and food preparation activities.47
Data processing and data collection

Data processing is the second category most readily automatable (69percent) and
accounts for 16percent of all the time spent working in the United States. That is followed
by data collection (64percent automation potential and 17percent of time spent). These
activities are common to almost all sectors, ranging from human resources staff recording
personnel history to mortgage brokers filling in forms, medical staff compiling patient
records, and accounting staff processing payments. These are not just entry-level or lowwage jobs; people whose annual incomes exceed $200,000 spend some 31percent of their
time doing those things as well.
47
Melia Robinson, This robot-powered burger joint could put fast food workers out of a job, Business Insider
UK, June 30, 2016.
43
Exhibit 9
Technical potential for automation across sectors varies depending on mix of activity types
Size of bubble indicates % of
time spent in US occupations
Ability to automate (%)

0
Sectors by
activity type
Manage Expertise
Interface
Unpredictable
Collect
physical
data
Process
data
Predictable
physical
50
100
%
Accommodation
and food services
73
Manufacturing
60
Agriculture
58
Transportation and
warehousing
57
53
Retail trade
51
Mining
Other services
49
Construction
47
Utilities
44
Wholesale trade
44
Finance and
insurance
43
Arts, entertainment,
and recreation
41
Real estate
40
Administrative
39
Health care and

social assistances
36
Information
36
Professionals
35
Management
35
Educational
services
27
44
Companies have long automated activities such as administering procurement, processing

payrolls, calculating material-resource needs, generating invoices, and using bar codes
to track flows of materials. But as technology progresses, computers are helping increase
the scale and quality of these activities. For example, a number of companies now offer
solutions that automate entering paper and PDF invoices into computer systems or even
processing loan applications. Robotic process automation systems use software to
automate well-defined data transactions currently performed by many workers.
Financial services and insurance provide one example of this phenomenon. The world of
finance relies on knowledge work, not physical work: stock traders and investment bankers
live off their wits. Yet about 50percent of the overall time of the workforce in finance and
insurance is devoted to collecting and processing data, where the potential for automation
is high. Insurance sales agents spend considerable time gathering customer or product
information, as do underwriters on verifying the accuracy of records. Securities and
financial sales agents prepare sales or other contracts. Bank tellers verify the accuracy of
financial data. As a result, the financial and insurance sector has the potential to automate
activities taking up 43percent of its workers time.
Performing physical activities and operating machinery in

unpredictable environments
In creating our high-level groupings of activities, we divided the performance of physical
activities and operation of machinery into two distinct categories depending on the
environment or setting in which the activity takes place. As we have seen, carrying out this
type of physical activity in a predictable setting has a rote quality to it, and the technical
potential to automate is accordingly high, at 81percent. When the environment or setting is
unpredictable, however, the automation potential is much lower, at 26percent.
Physical activities in an unpredictable environment make up a high proportion of the work
in sectors such as forestry and construction. Examples include operating a crane on a
construction site, providing medical care as a first responder, collecting trash in public
areas, setting up classroom materials and equipment, and making beds in hotel rooms.
The environment in these examples is not stable and can change in unpredictable ways.
Carrying out these activities thus requires a high degree of flexibility, which makes them
harder to automate. For example, people in public areas do not always drop trash in the
same place, and while sometimes they might drop paper, at other times they leave plastic
bottles or soda cans.
Interfacing with stakeholders

Across sectors of the economy, especially in services, workers interface on a regular
basis with a wide range of stakeholderscustomers, patrons, or visitors. Greeting them,
for example in a retail store, explaining technical product details or service information
to customers, including from a call center, responding to complaints and questions,
scheduling appointments, or providing advice are just some of the numerous types of
interactions that take can place with stakeholders. These types of activity for now have a
relatively low potential for automation based on currently demonstrated technologies, just
20percent. Social and emotional responses are important for these tasks, as are linguistic
and cognitive capabilities such as logical reasoning and problem solving.
45
Applying expertise to decision making, planning, and creative tasks

Even where computers perform above human levels in some well-defined activities such as
optimizing trucking routes, humansfor nowstill need to determine the goals, interpret
the results, or provide commonsense checks for solutions. Activities that require application
of decision making, planning, and creativity account for 14percent, about one-seventh, of
the total time spent working in the United States, but they have relatively low automation
potential, at just 18percent, based on adapting currently demonstrated technologies. These
activities can be as varied as coding software or creating a menu, developing marketing
plans, or writing promotional materials. They are common in fields such as education,
human resources, and finance, and considerable amounts of time spent on them in the
US economy involve evaluating students work, coordinating operational activities, and
examining financial records or processes.
Managing and developing others
The category of activities we describe as managing and developing others has the lowest
automation potential. Only about 7percent of time in the workplace is spent on these
activities, and the potential to automate them is low, about 9percent. Chief executive
officers and senior managers spend a significant proportion of their time engaged in such
activities, which brings down their overall automation potential; about 25percent of a CEOs
daily activities could be automated using currently demonstrated technology, but that mainly
represents data collection and analysis, rather than talent management.
Among the sectors, education is among the least susceptible to automation, at least
for now, with an automation potential of 27percent. To be sure, digital technology is
transforming the field, as can be seen from the myriad classes and learning vehicles
available online. Yet the essence of teaching is deep expertise and experience, and complex
interactions with other people. Together, those two categoriesthe least automatable of the
seven identified in Exhibit6account for about half of the activities in the education sector.
ASSESSING THE AUTOMATION POTENTIAL OF THE GLOBAL ECONOMY

We began our research into automation potential by focusing on sectors of the US economy,
to establish the methodological framework that underpins our research. We then broadened
the analysis to a total of 46 countries, using comparable national and international data,
where available, including wage data from foreign direct investment sources. Details of our
international methodology are in the technical appendix.
Overall, currently demonstrated automation technology has the potential to affect activities
associated with 40 to 55percent of global wages depending on country (Exhibit10). This
amounts to about $15.8trillion in wages and the equivalent of 1.1billion workers. As we
will see in Chapter 4, the actual deployment of automation could vary widely from country
to country, depending on a number of factors including the level of wages and the cost of
deploying solutions.
The key differences in the total wages associated with technical automation potential among
countries results from differences in the mix of sectors within each economy, the mix of
occupations within sectors, and wage levels.
In terms of total wages associated with technically automatable activities, the potential
is concentrated globally in China, India, Japan, the United States, and the five largest
European Union countriesFrance, Germany, Italy, Spain, and the United Kingdom. These
are the countries with a combination of the largest labor forces or higher wages.
46
Exhibit 10
The technical automation potential of the global economy is significant, although there is some variation
among countries
Employee weighted overall % of activities that can be automated
by adapting currently demonstrated technologies1
<45
4547 4749 4951
>51
No data
1 Pakistan, Bangladesh, Vietnam, and Iran are largest countries by population not included.
Automation potential is concentrated in China, India, Japan, the United States,

and the largest European Union nations
than half the wages and almost two-thirds of the total number of workers associated
SIMILARMore
from
ES (top of E4)
with technically automatable activities are in just four countriesChina, India, Japan,
and the United States. These four together account for about $9trillion of the wages and
more than 700million employees of the global total potentially affected. In the five largest
European Union economiesFrance, Germany, Italy, Spain, and the United Kingdom
more than 50million workers and $1.7trillion in wages are associated with technically
automatable activities (Exhibit11).
The largest amount of employment associated with technically automatable activities is in
China and India, because of the relative sizes of their labor force. Technically automatable
activities make up the equivalent of more than 600million full-time workers in the two
countries together. In terms of wages associated with technically automatable activities,
however, the United States is closer to Chinas level ($2.7trillion in the United States vs.
$4.1trillion in China) because of higher wage levels.
47
Exhibit 11
Technical automation potential is concentrated in countries with the largest populations and/or high wages
Potential impact due to automation, adapting currently demonstrated technology (46 countries)
Wages associated with
$ trillion
Labor associated with

Million FTE
%
Japan
United
States
Remaining
countries
55
Remaining
countries
2.7
5.1
100% =
$15.8 trillion
4.1
1.1
India
1.1
Japan
China
1.7
India
332
100% =
1,109 million
FTEs
35
Japan 54
60
Europe Big 51
United States
Europe Big 51
394
China
233
India
52
China
51
United States
46
Europe Big 5
46
Rest of world
50
1 France, Germany, Italy, Spain, and the United Kingdom.

Differences and similarities in automation potential globally

Our analysis of the technical automation potential of the global economy shows that there is
a range among countries of about 15 percentage points. Two factors explain this range. The
first is the sectoral makeup of each economy. That is, the proportion of a national economy
that is in sectors such as manufacturing or accommodation and food services, which both
have relatively high automation potential, compared with the proportion in sectors with
lower automation potential, such as education, management, and health care. The second
factor is the occupational makeup of sectors in different countries. That is, to what extent
workers in these sectors are engaged in job titles with high automation potential such as
manufacturing production, and those in job titles with a lower automation potential such
as management and administration. This weighting changes the automation potential of a
sector depending on the country.
Two examples illustrate these differences. The first is China and India (Exhibit12). With
the worlds largest workforces, they have similar automation potential and dynamics
overall: both have technical automation potential of 50percent. They have similar top
sectors, including agriculture, manufacturing, retail, construction, and transportation and
warehousing, and the automation potential within each sector is very similar. Manufacturing
and retail play a larger role in China than India, whereas agriculture accounts for a
significantly greater share of hours worked than in India than in China as a percentage
of the total. Within the sectors, the essential differences result from varying types of job
families. For example, India has more welders and sewing machine operators engaged
in manufacturing production than China, and both of these job families have a higher
automation potential than many other types of jobs, such as managing and developing
people, and specialized expert technicians. At the same time, India has a lower proportion
of jobs requiring interactions with stakeholders and managing and developing people,
activities with low automation potential.
48
Exhibit 12
Differences in technical automation potential in India and China are driven by
differences in sector and occupation mix
Sectoral makeup of China and India
%; million FTEs
773
Rest
Construction
25%
(193)
11%
(85)
Retail trade
18%
(139)
Manufacturing
19%
(147)
$6.70
$4.41
$5.82
41%
54%
8%
(36)
10%
(45)
15%
(68)
Agriculture
$4.39
$2.40
42%
$3.79
56%
$2.54
67%
Construction
and extraction
Maintenance
China
India
45%
15
47%
11
47%
11
46%
86%
36
$5.33
Production
64%
42
51%
(232)
27%
(209)
Automation potential (%)
Share of major occupational

groups in manufacturing sector
%
454
16%
(73)
Hourly wage per sector ($)
$3.03
$1.11
49%
38
59%
Transportation
49%
China
88%
45
64%
India

The second example highlights differences between Japan and the United States
(Exhibit13). Japan overall has an automation potential of 55percent of hours worked,
compared with 46percent in the United States. This difference is primarily due to a different
sectoral mix in the two economies, and within those sectors, a different weighting of jobs
with larger or smaller automation potential. For example, the automation potential of Japans
manufacturing sector is particularly high, at 71percent (compared with 60percent in the
United States). Japanese manufacturing has a slightly larger concentration of work hours in
production jobs (54percent of hours vs. 50percent) and office and administrative support
jobs (16percent vs. 9percent). Both of these job titles comprise activities with a relatively
high automation potential. By comparison, the United States has a higher proportion of work
hours in management, architecture, and engineering jobs, which have a lower automation
potential since they require application of specific expertise such as high-value engineering,
which computers and robots currently are not able to do. These differences outweigh
the higher level of wages in the United States than Japan, which affect the business case
for automation.
Similar differences exist among countries globally, for example, between Argentina and
Brazil, France and Germany, or Kenya, Nigeria, and South Africa. A detailed look at all 46
countries we have examined is available online.48
The data visualization can be found on the McKinsey Global Institute public site at tableau.com:
48
49
Exhibit 13
Japans manufacturing sector has a higher technical automation potential than
US manufacturing because of a different mix of occupations
Sectoral makeup of Japan and the United States
%; million FTEs
40%
(10)
Rest
$13.77
43%
(57)
Automation potential (%)
Share of major
occupational groups in
manufacturing sector
%
132
26
Hourly wage per sector ($)
Architecture
and engineering
Japan
United States
20%
7
$21.88
17%
Education
Administrative
support and
government
Health care
services
5%
(1)
11%
(3)
13%
(3)
14%
(4)
Retail trade
Manufacturing
17%
(4)
$15.40 38%
$14.46 50%
10%
(13)
$23.35 39%
$17.68 38%
13%
(17)
$21.67 27%
$14.32 55%
14%
(18)
$14.24 71%
27%
Management
Office and
administrative
25%
16
64%
61%
$23.38 26%
54
11%
(15)
$15.48 53%
9%
(12)
$22.05 60%
83%
Production
50
79%
United States
Japan
The technical potential to automate work activities in the global economy by adapting
currently demonstrated technologies is already close to 50percent, thanks to rapid
advances in automation technologies. That does not mean automation will occur overnight,
however, since technical potential is only one of several factors that will eventually lead to
automation adoption in the workplace. Only a tiny fraction of entire occupations could be
automated with current technologies, but activities across a wide spectrum are susceptible,
especially those involving predictable physical activities, data processing, and data
collection. What is the outlook for the development of these capabilities? How rapidly could
technical potential become workplace adoption? And what would the workplace then
look like? In the following chapters, we model some ways automation could arrive in the
workplace and lay out timeline scenarios for its actual adoption across the world.
50
51
Giovanni Mereghetti/MARKA/Alamy Stock Photo

52
3. FIVE CASE STUDIES

Our analysis of automation potential inevitably raises
questions about the practical effect on workplaces
and on the people who work in them. What exactly will
change, and how? To develop a vision of automation as
it could be applied in the workplace, we have created
hypothetical case studies across different industries that
suggest how automation could affect specific processes.
In this chapter, we outline five of these cases by way of
example. The five are from highly varied sectors and
involve a broad range of activities across various industry
sectors and physical settings: a hospital emergency
department, aircraft maintenance, an oil and gas control
room, a grocery store, and mortgage brokering.
For all their differences, our case studies also have
several essential elements in common. First is the
changing nature of work itself, which will likely affect all
workers at all skill levels. There will be less routine and
repetitive work based on rules-based activities, because
this can be automated across many occupations and
industries. This in turn will mean many workers may need
to acquire new skills. The workplace will become ever
more a place where humans and technology interact
productively. A major part of most human jobs will involve
working together with artificial intelligence, robotics,
and other technologies. Automation will affect more
than distinct work activities: processes and procedures
will also likely have to adapt, too. This in turn will have
profound implications for how the workplace is structured
and organized.
A second essential element of automation in the
workplace will be its impact on business economics.
Advances in automation technologies are often depicted
simplistically as robots replacing humans. In fact, in
evaluating the economics of our case studies, we see
that automations effects are twofold. While it will result
in some degree of labor substitution in all five of our
case studies, it will also drive significant performance
and quality gains in four of the five and, in one of
these casesthe oil and gas control roomthese
performance improvements far outweigh the gains from
labor substitution. These improvements are the result
of automated systems carrying out a range of activities
better than human workers, and include eliminating
waste, improving efficiencies, heightening safety, and
enhancing quality. As a rule of thumb, capital-intensive
industries tend to accrue automation benefits mostly

through performance gains, as capital is better utilized,
while labor-intensive sectors tend to benefit more from
labor substitution. Together they provide a strong
rationale for the deployment of automation technology.
The third essential element is that the business case
for automation is often strong. The relative cost of
automation is likely to be modest compared with the
value it can create. The types and sizes of investment
needed to automate will differ by industry and sector.
For example, industries with high capital intensity that
require substantial hardware solutions to automate and
are subject to heavy safety regulation will see longer lags
between the time of investment and the benefits than
sectors where automation will be mostly software-based
and less capital-intensive. For the former, this will mean a
longer journey to breakeven on automation investment.
However, our analysis suggests that the business case is
compelling regardless of the degree of capital intensity:
the run-rate benefits of investment in automation in our
case studies are between three and 11 times the costs of
that investment.
These case studies are not precise projections. The vision
of the future they provide is based on the hypotheses of
industry experts about how these processes could be
transformed by automation. While these case studies are
partly anecdotal, we nonetheless believe they are a useful
exercise that could indicate key benefits and challenges
related to developing and deploying automation
technologies in various sectors, as well as providing
a vision of how automation could actually transform
the workplace.
For all five cases, we outline two scenarios: an interim
future state as automation technology is deployed in
current processes and structures, and a provocative
future state in which we envisage automation being used
in a changed structure or process environment that is
tailored to maximize its advantages. We also outline
some of the likely barriers to automation deployment in
each of the cases, including legal and policy obstacles,
organizational impediments to change, difficulties
in integrating technology, economic viability, and
human reluctance to accept technology solutions in
certain situations.
IN A HOSPITAL EMERGENCY DEPARTMENT, LESS WAITING FOR

BETTER DIAGNOSES AND TREATMENT
Two common characteristics of many hospital emergency
departments today are their high level of human
interaction and long patient waiting times; in the United
States it is rare to be discharged in less than two hours
on average.49 Automation has the potential to reduce
those waits and increase productivity, as doctors and
nurses focus more effectively on better outcomes, and
machines take on routine activities such as registration,
checkout, and dispensing of prescriptions (see illustration,
Hypothetical future state of a highly automated
emergency department). Predictive health care using
sensing wearables to check vital medical signs and
remote diagnostics could cut patient waiting times. For
hospitals, automation could streamline billing and other
administrative activities.
To achieve such outcomes, hospitals will need to make
significant investment in automation technology, along
with time and capital to train staff. They will also need to
redesign process workflow. Doctors and nurses will have
to become comfortable working closely with and trusting
automated systems. Safety and liability are significant
challenges in a sector where malpractice suits are
common; in the United States, about $3.6billion was paid
out in malpractice suits in 2014, and artificial intelligence
companies could find themselves on the receiving end.50
Stringent privacy regulations will need to be safeguarded.
The emergency room is also a place where human
unease with machines could be strong: people who
come to hospitals for medical emergencies usually want
and need to interact with qualified humans and may
not feel comfortable with machines, however medically
competent.51
In emergency departments in the United States today,
about 80 to 85percent of the patients are walk-ins, and
about the same percentage of the total are treated and
sent home usually with prescribed medicines.52 Patients
interact with a range of workers. First are the medical
secretaries who enter and validate patient information.
Triage nurses check vital signs, request laboratory tests

or imaging, and decide to discharge or refer a patient
to see a doctor. Doctors examining a patient prescribe
medicines and decide to discharge, admit to hospital, or
refer the patient to a specialist. Lab technicians conduct
tests. At the end of the process, medical secretaries
collect payment or compile documentation for an
insurance claim.
At least some of these activities could be fully or partially
automated. They include the initial work of gathering
patients information, checking vital signs and requesting
lab reports. Lab registration and tests are also potentially
automatable, as is the end process of payment. Some
aspects of a doctors work in an emergency department
are also potentially automatablenot just the data
collection that takes up some of a doctors time, but
also some areas of disease diagnosis, and even some
aspects of medical procedures and surgery. For example,
in radiology, computers are already analyzing X-rays,
CT scans, and MRI imagery. Completely automated
diagnosis is not likely to happen quickly, partly for reasons
of patient acceptance, and partly because of the technical
difficulty of integrating data from multiple sources
(including natural language understanding, recognizing
and processing the patients emotions) to determine a
diagnosis and course of treatment. Automated diagnostic
advice is thus likely to augment doctors decision making
before fully automated diagnosis, except perhaps in
special instances, such as radiology, or cell pathology
(checking for abnormalities through a microscope).
Overall, we calculate that about 30percent of the benefits
of automation in an emergency room would come
from performance gains, and 70percent from labor
substitution. Productivity could rise substantially, while
the number of full-time equivalents could decline by about
half, with the main reductions at the registration desk and
in lab testing.
ProPublica lists average waiting and treatment times in US emergency rooms, broken down by state, at https://projects.propublica.org/
emergency/. See also Lisa Esposito, Enduring really long waits at the emergency room, US News and World Report, May 8, 2015.
50
2015 Medical malpractice payout analysis, Diederich Healthcare, May 2015.
51
The unsettling feeling that people may experience when humanoid robots or audiovisual simulations closely resemble humans is known as
uncanny valley, a term coined by a Japanese robotics professor, Masahiro Mori, in a 1970 essay. The eeriness is largely prompted by the
machines appearance, which is realistic but not convincingly so.
52
Cheryl Gutherz and Shira Baron, Why patients with primary care physicians use the emergency department for non-emergency care, Einstein
Journal of Biology and Medicine, volume 18, number 4, 2001.
49
54
3. Five case studies
Hypothetical future state of a highly automated

emergency department
Triage nurses would
be aided by automated
fast diagnostics
using blood and
auto-generated reports
on basis of vitals and
tests
Patients pre-register
by mobile phone.
On arrival, they are
issued a wearable
monitoring device
that collects vitals
Fully automated
checkout including
medicines, billing
and issuing reports,
or, in the case of
hospitalization, bed
assignment
Lab tests would be
automated, including
report generation, for
improved accuracy
Potential economic
benefits of automation
70% Labor substitution
Autonomous tugs
can pull beds and
bring medicines and
instruments to the
point of care. Drugs
are dispensed by
automated pharmacy
AI diagnoses and
advice on complex
and high acuity
cases contribute to
better outcomes
30% Performance gain

Performance gains
Algorithms
recommend
diagnosis and
treatment to
doctors and nurse
practitioners
Increased productivity
of nurses and doctors
Reduced patient
waiting time
Better health care
outcomes
11% Relative impact1
1 Ratio between additional net
impact and operating cost.
55
FOR AIRCRAFT MAINTENANCE, AUTOMATION WILL RAISE SAFETY,

IMPROVE DEFECT DETECTION, AND REDUCE TIME WASTED ON
WALKING AND WAITING
Maintaining a commercial aircraft takes an average
of about 8,000 technician hours per plane annually.53
Technicians conduct visual inspections of the aircraft for
signs of physical wear and damage. They also remove
and replace parts, conduct a sign-off for quality, and
handle administrative tasks, including supplying records
to the Federal Aviation Administration where needed.
These are jobs that involve spending considerable
amounts of time on walking around the aircraft and
waiting on parts, planes, or people. They are also
dangerous: in 2013, 57 job-related deaths in the United
States were in aircraft maintenance.54
Automation could have a major beneficial impact on
the sector (see illustration, Hypothetical future state of
highly automated aircraft maintenance). First, removing
technicians from fall hazards and fuel tanks would
represent a significant improvement in safety. Second,
robots equipped with image process algorithms already
do a better job than humans at image identification, and
deploying them would improve the detection of defects in
the aircraft; maintenance errors caused more than onethird of the 179 commercial jet engine accidents between
1988 and 2013.55 Automated warehouse systems
could eliminate about 75percent of the time wasted by
walking around the aircraft, picking up tools and parts,
and improved sensors and analytics could raise the
proportion of predictive maintenance, which is less costly
than reactive maintenance. For maintenance companies,
the savings from reduced waste and a move to on-
demand maintenance would save costs. Automation

could enable experts to monitor all maintenance from a
command center, which would reduce variability, and
ensure better data collection. Consumers would benefit
from these savings if they are passed on in the form of
lower flight costs. About half of engine-related delays
today are caused by maintenance issues, and so flight
delays could also be reduced.56
For technicians themselves, automation could change the
workplace and their roles. Much of the routine work they
carry out, including walking around the plane, moving it,
and logging work records, could be automated. Already in
an interim phase, remotely controlled robots could crawl
through planes and inspect fuel tanks. High-resolution
cameras guided by experts could inspect exteriors.
Artificial intelligence algorithms could suggest potential
problems based on logs even before inspection takes
place. In our more advanced automation scenario,
small robots could inspect the airframe without moving
panels. Automated tugs would move planes, while
robotic carts bring and remove parts and tools based
on scheduling routines. These and other changes would
allow technicians to be more focused on knowledge and
handling exceptions, which require greater training.
Overall, we estimate that 35percent of the value created
by automation in aircraft maintenance could come from
performance gains, while 65percent could potentially
come from labor substitution.
Stephen Holloway, Straight and level: Practical airline economics, Routledge, 1997.
US Bureau of Labor Statistics, 2015.
55
Mary S. Reveley et al., Causal factors and adverse events of aviation accidents and incidents related to integrated vehicle health management,
National Aeronautics and Space Administration, March 2011.
56
Jeffrey OBrien, Maintenance: What causes aircraft accidents? Fiix, May 9, 2012.
53
54
56
Hypothetical future state of highly automated

aircraft maintenance
Machine vision to
identify many
common defects.
Predictive
maintenance is far
less costly than
reactive maintenance
Instead of humans
searching for wear and
tear visually, drones
inspect exterior, and
tiny insect-inspired robots
check airframe without
removal of panels
In place of logging
inspection status
manually, compliance
documents are
automatically routed
to a secure site for
easy regulator access
Fewer technicians
remain on the
maintenance hangar
floor, but they spend
more time on problem
solving, and require
more continual training
Diagnostic
algorithms and
inspection robots
record all their
own compliance
information
automatically
Potential economic
Robots do physical
tasks. Humans can
avoid danger zones,
and parts are installed
faster and with less
variability
Automated warehouse is
linked to procurement
department with robotic
part and tool delivery.
Automatic ordering
based on forecast
demand raises efficiency
Automated tugs
rather than technicians
move plane in and out
of hangar. Less human
time spent on tasks that
could be done by
automation cuts waste

Performance gains
Improved safety
Better defect detection
Predictive maintenance
Elimination of time
wasting

57
IN AN OIL AND GAS CONTROL ROOM, PREDICTIVE MAINTENANCE

AND OTHER ADVANCES COULD BRING SIGNIFICANT
PERFORMANCE GAINS
Most of the gains of automation in oil and gas control
rooms are likely to come from higher productivity and
safety (see illustration, Hypothetical future state of
automation in oil and gas operations). A control room is
an operations center that monitors and controls upstream
exploration and drilling operations. For now, key roles in
the oil and gas sector are divided between onshore and
offshore facilities, with operations and maintenance on an
offshore rig largely overseen and carried out by managers
and workers on the rig, while teams of petroleum and
other engineers provide technical support and coordinate
activities from headquarters onshore. Offshore work
can be a dangerous business: the job-related fatality
rate of people working in extraction operations in 2013
was four times the average for the US economy.57 The
performance and training of offshore operators vary,
and errors are sometimes made, including ones with
severe consequences.
Automation technologies can considerably raise the
performance of control rooms by removing operators
from environments that are hazardous and expensive
to maintain and by capturing data that can be used for
predictive and preventive maintenance and operational
best practice. Centralizing expertise in an onshore
location can improve strategy planning and the
effectiveness of event responses. In our more futuristic

scenario, permanent seafloor robots could undertake
repairs or conduct additive manufacturing, while 3D
printers on a surface vessel could print out replacement
parts as needed. Robots would conduct standard
maintenance, and operating algorithms developed using
historical logs could deliver more efficient operations and
greater safety.
The advantages of automation in this case include better
personnel safety, greater efficiency, higher throughput,
improved agility, and cost reductions from relocating
operators from remote sites to centralized offices.
Improved sensors for remote operations and analytics
can enable predictive maintenance, which is just onequarter the cost of reactive maintenance.58
Overall, we estimate that 80percent of the value created
from automation in oil and gas control rooms would
come from performance gains, with the rest from labor
substitution. To reap such gains will require the integration
of technologies so that cognitive, sensory perception,
and physical capabilities can be deployed on site. Data
scientists, engineers, and developers will be needed to
develop algorithms that optimize remote management.
Occupational fatalities during the oil and gas boomUnited States, 20032013, Morbidity and Mortality Weekly Report, Centers for Disease
Control and Prevention, May 29, 2015.
58
Elly Earls, Staying one step ahead with predictive maintenance technology, Mining-technology.com, October 15, 2013.
57
58
Hypothetical future state of automation in oil

and gas operations
With fewer offshore
people to
coordinate, onshore
operations and
maintenance are
streamlined
Robots perform
predictive maintenance offshore and
human operators
are reduced from
the hazardous
environment
Petroleum engineers
are now joined by
programmers
onshore who work
to optimize
throughput
Engineers onshore
focus on innovation
and efficiency,
designing modular
equipment with
maintenance routines
for robots
Potential economic
Offshore surface
vessel 3D prints
components of
subsea factories
for robot vehicle
installation to
equipment
Subsea factories
act autonomously
for standard and
safety-oriented
operations
controlling flow and
limiting accidents
Offshore seafloor
robots operate,
modify, and install
small parts in
dangerous areas
previously serviced
by human divers

Performance gains
Better safety, as
people removed
Faster decision-making
through global control
room
Predictive maintenance
saves costs
59
AT A GROCERY STORE, IMPROVED CUSTOMER EXPERIENCE

AND HIGHER PRODUCTIVITY, AS THE NEED FOR HUMAN
STAFF DECLINES
Imagine walking into a grocery store and being greeted
by name, courtesy of facial recognition software. Thus
begins a highly personalized shopping experience that
is faster, more tailored to your preferences, and more
convenient than shopping today. If you cannot find what
you are looking for, you provide instant feedback. You
might use personalized coupons on your mobile device.
Once you have finished your selection, an automated
back-room service sends out your goods, or a drone
drops them off at your home. Best of all, there is no line
for payment because there is no physical checkout. The
store senses what goods you have with you when you
walk out of the store, and payment is accurately and
automatically deducted from your preferred account (see
illustration, Hypothetical future state of automation in a
grocery store).
We are not all that far from such a scenario today. Stores
still have physical shelves, but self-checkout kiosks are
becoming commonplace and are a short step from
automatic payment and shipping.59 Robot cleaners and
automated storerooms already exist. Augmented virtual
reality product views are just a question of time, where
consumers will be able to look at displays of goods
(probably using special glasses or other technology) to
find out more information about the product such as
ingredients and nutritional details. One of the biggest
overall benefits of automation in retail will be the improved
customer experience, as the online and offline shopping
experience merges into one, even as it becomes
more individualized.
From the employment perspective, our case study
suggests that automation in grocery stores could have a
significant impact on staffing needs, with a reduction of
about 65percent of hours, mainly for front-end cashiers
and people engaged in stocking and cleaning. Some
59
workers could be repurposed toward higher value-added

activities such as customer engagement.
The retail sector overall will face some significant
changes. Space productivity will rise, and that in turn
will reduce the need for large stores. Smaller stores
require less investment; we estimate savings of 60 to
80percent. Lower inventory and working capital will also
be a feature of the new retail business landscape, as the
brick-and-click models merge into a seamless whole, and
retailers leverage physical stores as distribution centers.
Data analytics will enable retailers and manufacturers to
customize and target promotions.
Some customers may find marketing use of their personal
data, including location-based alerts, overly intrusive.
Many jobs in retail are entry-level, low-skill ones, and
the elimination of many such positions could cause a
public outcry. Training will be essential so that staff can
troubleshoot technical and other problems and make
recommendations. From a technology perspective,
sensory perception will need to be integrated with pattern
matching, so that customers are recognized and given
relevant recommendations.
While the potential savings from smaller formats, less
inventory, and lower payroll could be significant, the slim
margins in retailing may mean that only large chains
will have the capital needed to make the investment in
full automation. Overall, we estimate that the benefits
of automation will be three times the cost. Along with
health care, this is the lowest ratio of our five case studies.
Labor substitution gains in retail accounts for 68percent
of the potential gains, compared with 32percent from
improved performance.
Amazon in December 2016 launched a store without cashiers, Amazon Go. Leena Rao, Amazon Go debuts as a new grocery store without
checkout lines, Fortune, December 5, 2016.
60
Hypothetical future state of automation

in a grocery store
No lines and no-stop
checkout, thanks to
automatic payment
and shipping. Just
head home, and the
groceries will follow
right behind you
Continuous
inventory
tracking through
sensors and video
Automatic facial
recognition and
personalized
greeting on arrival.
Virtual assistants
provide directions
Automated stock
room manages
inventory and
prepares custom
orders, with delivery
by autonomous
drone or vehicle
Potential economic
The footprint of the

average US supermarket
is the size of two football
fields. Automated stores,
without the need to
display so much of their
wares, would be smaller
Purchases are made

by mobile phone.
Customers receive
personalized
mobile coupons,
and can give instant
feedback
An automated shopping
cart follows you around.
The shelves are
monitored and
restocked
automatically by
robots

Performance gains
Improved customer
experience
Higher space
productivity through
backroom automation
Lower inventory and
working capital needs
61
MORTGAGE APPLICATIONS: SPEEDING UP THE

PROCESS, REDUCING DROPOUT RATES, AND RAISING
CUSTOMER SATISFACTION
The financial-service industry has been at the forefront
of adopting automation technologies for a range of
back-office work. This case shows how that degree of
automation could reach even higher levels. It focuses
on the business of processing mortgage applications,
which consists in large part of collecting and processing
data, two categories of activity that already have a high
automation potential, by adapting currently demonstrated
technologies. Our scenarios for automation in this
process accordingly correspond to a sharp reduction
in the overall required labor hours, of between 55 and
85percent. The speed with which mortgages are
processed could accelerate substantially. For now, it
takes an average of 37 days to approve a mortgage
application in the United States, of which about 14 to
21 days are spent on the mechanics of application
processing.60 In our interim automation scenario, in
which technology is deployed in current processes and
structures, this could drop to less than six days. In a more
futuristic scenario, in which the processes themselves
change, mortgage approvals could come through in less
than a day. The shorter turnaround time could improve
the dropout rate by possibly 30percent or more, as many
people who apply for mortgages today drop out as a
result of the lengthy process.
For mortgage companies, the improvements derived
from automated processes could lower default risks
and eliminate inconsistencies in processing, thereby
reducing the need for human performance management.
Automation could also create a potential for partnership
between mortgage companies and real estate agents for
real-time lead generation, mortgage application, and loan
fulfillment. Customer satisfaction could rise as a result
of instant pre-approval, hassle-free applications, and
much faster turnaround. Moreover, some of the industrys
challenges, including human bias in underwriting and
the difficulty that underserved borrowers can have in
accessing affordable capital, could also be eased by
automated processes.
60
61
Of our five cases, however, the automation of mortgage

applications will potentially weigh most heavily on labor.
The vast majority of the impact in this process, about
88percent, would come from labor substitution gains,
compared with just 12percent from performance gains.
Moreover, the business case to move to automation is
strong, since the software costs could be relatively low,
and wages of loan officers are relatively high, at $35 per
hour.61 The run-rate benefit from automated mortgage
origination could be as much as 11 times the cost of
automation itself, according to our estimates, easily the
highest ratio of our five case studies.
For mortgage brokers, automating and speeding up the
approval processes will allow more time on complex
tasks such as advising customers or handling exceptions
that require human expertise and judgment, as well as
additional time to manage unusual applications. That
in turn will require more training. Data scientists will
also be needed to increase the accuracy of algorithms
used in processing, and to integrate platforms and data
sources. Nonetheless, there is likely to be considerable
restructuring and elimination of redundant positions in
the core application processing functions, which could
require managing labor issues, including helping to
redeploy displaced workers.
To reap the full benefits of automation, the entire
mortgage approval process will need to be redesigned
and piloted in select branches to ensure all necessary
resources are available. There are technological
challenges in ensuring end-to-end integration with
systems from retail banks and underwriters, such as
handling documents, messaging, and conducting risk
analysis. However, given the potential cost savings and
fairly certain return, financial institutions have an incentive
to invest in automation for their mortgage origination
activities, much as they have for straight-through
processing of other types of transactions.
Evan Nemeroff, Survey shows average mortgage closing time hits three-year low, National Mortgage News, August 22, 2014.
Occupational Employment Statistics, US Bureau of Labor Statistics, 2014.
62

No sector of the economy will be immune to automation. While our five hypothetical case
studies are not precise predictions about how automation will be deployed in the workplace,
they do signal some of the key benefits and challenges to come. Integrating technological
capabilities, retraining staff, adjusting processes, and, in some sectors, working out how to
make workers and customers comfortable interacting with machines rather than humans,
are among the considerable changes that will be required to maximize the benefits from
automation. Gains from labor substitution could be significant, but they will vary from
sector to sector. Performance gains, in the form of better quality, greater safety, and higher
productivity, could also be substantial, to judge by the findings of our case studies. How
quickly will any of this happen? In the next chapter, we analyze the factors that will hasten
or slow the adoption of automation, and project timelines for its implementation in the
global economy.
63
Fully automated vehicles moving containers at the port of Los Angeles.

Tim Rue/Bloomberg/Getty Images
64
4. FACTORS AFFECTING THE PACE

AND EXTENT OF AUTOMATION
Automation will be a global force, affecting all countries, all sectors, all jobs, and all work
activities. Already today, machines and algorithms are playing a much larger role in the
workplace, but how soon will it be before we all feel the impact of automation technologies?
Could machines really carry out much or most of the work humans do todayand if so,
by when?
In this chapter, we describe factors that can affect the pace and extent to which automation
is adopted in the global economy. We also present the results of a model that includes
a range of potential scenarios illustrating how the automation of existing work activities
could evolve.
FIVE FACTORS INFLUENCING AUTOMATION OF WORK ACTIVITIES

Overall, we have identified five broad factors that can influence the pace and extent
of automation of work activities. They are: technical feasibility; the cost of developing
and deploying solutions; labor market dynamics; economic benefits; and social and
regulatory acceptance.
Technical feasibility
Technology has been automating human activities for centuries, from the printing press
to the steam engine and the internet, and fundamentally reshaping the economy in
the process. Over the past two centuries, the share of people working on the land in
advanced economies has fallen from a majority to a tiny fraction. More recently, the United
States and other advanced economies have seen a decline in the share of the workforce
engaged in manufacturing. Technological advances require basic scientific research, but
in order for these advances to be adopted, they also require engineering solutions, or
applied research. Both take time to develop. There is a lag between a technology being
demonstrated, and a viable product being developed using that technology. Orville and
Wilbur Wright pioneered flying an aircraft in 1903, for example, but it took 11 more years
before the first commercial flight, across Tampa Bay, in Florida, took place, and the true birth
of commercial aviation in the United States is usually dated back to 1926, when pioneering
operators had to begin complying with federal regulations.62 There was a similar lag in the
development of automobiles, between German engineers Nikolaus Otto, Gottlieb Daimler,
and Wilhelm Maybach, patenting the compressed charge, four-cycle engine in the 1870s
and production of automobiles on a commercial scale some 15-20years later. Solutions
have to be engineered for specific use cases. While from a scientific perspective, a light
passenger vehicle is conceptually the same as a tractor-trailer, the engineering required to
develop these two solutions has very different specifications and each requires time and
energy. Similarly, while you could view predictive maintenance of a complex engineered
system such as a power plant and preventive health care for a congestive heart failure
patient (predictive maintenance of the human system) as being conceptually the same
problem, actually creating the software and models to prevent a failure in these two systems
require quite distinct and considerable engineering efforts.
At times, a paradigm shift is needed in how a new technology should be applied. When
steam engines in factories began replacing the water wheel, everything was driven from a
central mechanical drive. The first attempts at using electricity tried to duplicate this, but
Ronald E.G. Davis, The birth of commercial aviation in the United States, Revue belge de philologie et
dhistoire, volume 78, number 3, 2000.
62
only really made a difference when people realized it would be more efficient to distribute
electrical energy through the building, and have separate electric motors in individual
machines, rather than one central motor that distributes mechanical energy.63
Today, technological advances in automation abound in areas from physical hardware and
robotics to artificial intelligence and software, as we discussed in Chapter 1. Yet much of
the innovation that we are seeing, from self-driving cars to digital personal assistants such
as Apples Siri, Amazons Alexa, and Google Assistant, are still in developmentand often
imperfect. That means there remains a lot of work to be done by scientists and engineers.64
Cost of developing and deploying solutions

The cost of automation technologies will affect when and where they are deployed.
The costs of developing and deploying solutions have to be recouped. Even when the
technology is purchased from a supplier, the supplier will amortize the development costs
into the pricing.
Developing and engineering automation technologies takes capital. Some technologies
require substantial physical infrastructure such as tooling and laboratories. But even
virtual solutions that are based on software require real investments in engineers to create
solutions. A decade ago, the largest corporate research and development spending was
to be found in automotive and pharmaceutical companies; today, technology companies
dominate, with companies including Amazon, Alphabet, Intel, and Microsoft spending
more than $10billion apiece on annual R&D.65 Hiring talent can also be costly. Google, for
example, acquired DeepMind Technologies in 2014, at an estimated price of $500million.
With approximately 75 DeepMind employees at the time of the deal, the price tag was nearly
$7million per employee. This is in line with other estimates by experts, who say that aquihires of cutting-edge AI startups cost around $5million to $10million.66
Deploying automation technologies also incurs costs. For physical technologies, these
are real capital expenditures. Industrial robots cost from tens of thousands to millions of
dollars. Replacing an ordinary heavy truck with a self-driving truck requires an expenditure
of capital. And the self-driving truck will likely be more expensive than the truck it replaces, at
least when the technology is first released.
These deployment costs are lower for software-based solutions, especially when
delivered remotely through the cloud, and where the software is sold as a service, thereby
turning capital expenditure into operating expenditure. But the cloud has a real physical
instantiation as data centers and networks, and they in turn represent costs. Moreover,
deployment of a software (or hardware) solution can almost never be done without
significant implementation services, such as the costs of customizing the software for an
individual organization, changing the processes within an organization, and training staff.
For enterprise software, the associated implementation services often cost several times
the costs of the software itself. All these costs affect the business cases for where and when
automation is adopted.
Nicholas Carr, The big switch: Rewiring the world from Edison to Google, W. W. Norton, 2008.
For recent research on the modeling of enterprise 2.0 adoption, see Jacques Bughin, Taking the measure of
the networked enterprise, McKinsey Quarterly, October 2015, and Martin Harrysson, Detlef Schoder, and
Asin Tavakoli, The evolution of social technologies, McKinsey Quarterly, June 2016.
65
Justin Fox, The big spenders on R&D, Bloomberg News, April 29, 2016.
66
The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016.
63
64
66
4. Factors affecting the pace and extent of automation
Labor market dynamics

The labor costs associated with work activities that could potentially be automated are
another factor that will affect the pace and extent of automation. These costs are affected by
the complex dynamics of labor markets.
Even for the same activity, for example, entering data into a financial system, there is a wide
range of wages paid, across different positionsfrom accounting clerks to chief financial
officersand in different companies, for example a small, family-owned business compared
with a Global 50 corporation. Wage rates also vary by geography.
Labor supply is determined by demographics, with the share of the working-age population
potentially declining in many countries in coming decades. But it also varies in terms of
skills, which are affected both by intrinsic talents of individual human beings, as well as by
education and training that people receive. People can learn new skills but it takes time and
money (see Box3, Technological change and skills)
Labor markets are dynamic systems. Supply, demand, and wages all vary over time. If
automation frees up human capital, then supply will increase, which could be redeployed
into other positions if the demand exists. But there could be a skills mismatch, which
will require time and training, delaying redeployment. There could also be information
asymmetries, which digital labor markets could help address.67
Box3. Technological change and skills

Economists who study the impact of technological
innovation on the workforce have noted varying effects
on workers at different skill levels, depending on the time
period.1 In the 19th century, for example, technological
changes raised the productivity of lower-skill workers and
created new opportunities for them, at times replacing
the craftsmanship of higher-skill artisans. This so-called
unskill-biased technical change reduced the value
of some high-skill workers even as it boosted lowerskill ones.
With the advent of information technology and the
internet, the reverse has happened: the productivity of
higher-skill workers, especially those engaged in abstract
thinking, or with creative and problem-solving skills,
has increased, while the relative demand for lowerskill workers has not. This phenomenon of skill-biased
technical change is manifested in a number of ways. In
advanced economies, median income households have
been receiving a lower share of the total wage share of

GDP, in part because demand for less-skilled workers
has dropped, even as demand for high-skill labor has
risen.2 In 1981, college-educated workers in the United
States earned a 48percent wage premium over high
school graduates. By 2005, that premium had risen
to 97percentin other words, an American college
graduate earns almost twice as much as a high school
graduate.3
What will the spread of automation mean for workers at
different skill levels? Erik Brynjolfsson and Andrew McAfee
discuss a new shift to talent-biased technical change
which has created very high demand for superstars such
as experts in artificial intelligence or data scientists.4 Our
analysis suggests that all workers at all skill levels have
the potential to be affected at least partially by automation
based on currently demonstrated technologies.
See David H. Autor, Frank Levy, and Richard J. Murnane, The skill content of recent technological change: An empirical explanation, Quarterly
Journal of Economics, November 2003, and Daron Acemoglu and David H. Autor, Skills, tasks, and technologies: Implications for employment
and earnings, in Handbook of Labor Economics, volume 4B, David Card and Orley Ashenfelter, eds., Elsevier, 2011.
2
See, for example, Global wage report 2012/13, ILO, December 2012; OECD economic outlook 2012 volume 1, OECD, June 2012; Andreas
Hornstein, Per Krusell, and Giovanni L. Violante, The effects of technical change on labor market inequalities, Center for Economic Policy
Studies working paper number 113, July 2005.
3
David Autor, Skills, education, and the rise of earnings inequality among the other 99percent, Science, volume 344, issue 6186, May 2014.
4
Erik Brynjolfsson and Andrew McAfee, The second machine age: Work, progress, and prosperity in a time of brilliant technologies, W. W. Norton
67
See A labor market that works: Connecting talent with opportunity in the digital age, McKinsey Global Institute,
& Company, 2014.
June 2015, and Independent work: Choice, necessity, and the gig economy, McKinsey Global Institute,
October 2016.
1
67
Demand is also not static. New types of work activities and jobs are created all the time,
even as new technologies come on the market. Bank tellers and ATMs are one example.
ATMs were first installed in the United States and other developed economies in the 1970s,
and from the mid-1990s, banks rapidly increased their use of ATMs. Contrary to some
expectations, their advent actually increased the demand for human tellers, for two reasons.
First, ATMs reduced the cost of operating a bank branch, and banks responded by opening
more branches. Fewer tellers were needed in each branch, but more branches meant
that teller jobs did not disappear. Second, the tasks that ATMs could not doin particular,
developing relationships with clients such as small businessesbecame more valuable. For
tellers, the nature of their activities changed, with cash handling becoming less important
and human interaction more important.68
The relative costs of labor vs. automation will affect the pace and extent of adoption. If
workers are in abundant supply and significantly less expensive than automation, this could
be a decisive argument against it. For example, food service is one activity with a high
automation potential based on adapting currently available technologies. However, current
wage rates for this activity are among the lowest in the United States, reflecting both the
skills required and the size of the available labor supply. Since restaurant employees who
cook earn an average of about $10 per hour, a business case based solely on reducing
labor costs may be unconvincing.
Economic benefits
The potential economic benefits from automation are not limited to labor cost reductions.
As we noted in the case studies in Chapter 3, performance gains include increased profit,
increased throughput and productivity, improved safety, and higher quality, which are harder
to quantify than labor costs but no less tangible. In our hypothetical look at automation
of an oil and gas control room, for example, performance gains accounted for more than
three-quarters of the total benefits. There are also indirect benefits, such as wage growth,
and automations potential to create business and economic incentives that drive corporate
decision making, and unleash entrepreneurial energy. Automation could also spur policy
makers. However, these indirect benefits will have to be netted out against indirect costs
caused by automation, such as those associated with labor displacement
Regulatory and social acceptance
Automation faces some significant regulatory and social barriers to implementation. They
include safety and liability issues. One accident could trigger stringent regulations. Artificial
intelligence used in military robots or autonomous vehicles may have to make judgments
that harm people, creating moral controversy. Technology makers could also be exposed to
legal product liability if robots malfunction.
Privacy is also a potential barrier, especially in areas where personal data is highly sensitive,
such as health care; already, health care IT can struggle to link together different data sets
due to mandatory anonymization of data.
From a social perspective, too, automation will need to overcome some barriers. If many
workers lose jobs and are unable to find new ones, the social and political pressures against
automation could become significant. Humans may not want to adopt new technologies or
work with automated products due to fears about job security.
Finally, personal preferences and discomfort with technologies could prevent automation in
settings where human relationships are important, such as for caregivers.
68
68
James Bessen, Toil and technology, Finance & Development, volume 52, number 1, March 2015.
MODELING SCENARIOS FOR POTENTIAL PACE AND EXTENT OF ADOPTION

To analyze a range of potential scenarios for the pace at which automation could affect
activities across the global economy, we constructed a model that synthesizes the effects
of these five factors into four timing stages. We estimate when automation technologies will
reach each level of performance across 18 capabilities, the time required to integrate these
capabilities into solutions tailored for specific activities, when economic feasibility makes
automation attractive, and the time required for adoption and deployment (Exhibit14).
We modeled scenarios incorporating these stages for each individual activity in every
occupation for all sectors across 46 countries that account for about 80percent of the
worlds workforce.
Exhibit 14
Five factors affect the pace and extent of automation; we model using four stages
Key factor
Cost of
developing and
deploying
Labor market
dynamics
Economic
benefits
Regulatory
and social
acceptance
Impact on
pace and
extent of
automation
For an
activity to be
automated,
every
capability
utilized for
that activity
must reach
the required
level of
performance
Costs
associated with
developing as
well as
deploying
different
solutions
determine the
pace of reaching
economic
feasibility
Economic
feasibility of
automation will
depend on
comparison with
cost of human
labor, affected
by supply and
demand
dynamics
In addition to
labor cost
savings,
automation
could bring
more benefits to
employers,
including
increased
quality and
efficiency and
decreased error
rate
Adoption of
automation
shaped by pace
of organizational
change, policy
choices, and
acceptance to
stakeholders
Stage
How we
model it
Capabilities
need to be
integrated to
form
solutions
potential
Solution development
Economic feasibility
Adoption
Estimate the
Estimate solution
Assume adoption
Model an S-shaped
technology progression
timeline for each
capability through
interviews and surveys
with industry and
academic experts
development times for

activities based on
required capabilities
and historical
development timelines
begins when
automation cost for an
activity is at parity with
labor cost
Compare labor wage
and solution cost
Occupation- and
country-specific
wages and their
evolution
Solution-specific
costs and their
reduction
adoption curve based

on historical technology
adoption rates
NOTE: Economic benefits affect both when adoption will begin and its pace. For determining economic feasibility, we assume that decision-makers discount the
uncertain benefits of initial labor cost savings by roughly the same amount as they believe the also uncertain non-labor cost-related benefits will be captured.
69
The scenarios we have modeled create a time range for the potential pace of automating
current work activities. We have created two theoretical bookends, an earliest scenario,
in which all of the modeling parameters are flexed to the extremes of the set of plausible
assumptions that would result in faster automation development and adoption, and a
latest scenario, in which we flex all of the parameters in the opposite direction. The reality
will likely fall somewhere between the two. Modeling all of these factors, the date at which
50percent of the worlds current work activities are automated could be around 2055, but
we posit possible scenarios where that level of adoption occurs up to almost 20years earlier
or later. (Exhibit15).
Exhibit 15
Automation will be a global force, but adoption will take decades and there is significant uncertainty on timing
Time spent on current work activities1
%
Adoption
Early scenario
Late scenario
potential
Early scenario
Late scenario
100
90
80
70
60
50
40
30
20
10
0
2016 20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
2100
1 Forty-six countries used in this calculation, representing about 80% of global labor force.
We stress that we are not making specific point predictions; we use scenarios to describe
the envelope of potential outcomes, and we fully acknowledge the simplification that comes
with any modeling exercise. For example, we do not account for all of the dynamics of the
labor market described above, including whether wages for specific occupations would
decline because displaced workers would increase labor supply.
DUPLICATE from ES
The model is likely to be precisely wrong but we hope it is directionally right with regard
to overall findings. Occupations within sectors with high automation potential today, jobs
that involve types of activities that we categorize as easiest to automate, such as physical
activities in a predictable environment or data collection and processing, will most likely
be among the first to feel the impact of automation. From a geographical perspective,
advanced economies are likely to deploy automation ahead of many emerging economies,
because of higher wage levels, which make a stronger business case for deployment, as
well as the nature of the solutions that will be needed to integrate the technologies into
the workplace.
70
Furthermore, while the macro effects on a nations entire economy may be quite slow (we
model the adoption of automation across several decades), the micro impact on particular
work activities, in certain sectors, and in individual countries, could be quite fast, as
competitive pressures come together with advances in technological development. That
is, for an individual worker who is displaced by automation, or a company whose industry is
disrupted by competitors using automation to shift the basis of competition, these effects
could occur quickly indeed.
DEVELOPING AUTOMATION CAPABILITIES IS A KEY INITIAL DETERMINANT OF

ADOPTION TIMELINES
The deployment of automation in the workplace can begin only when machines have
the capabilities required to carry out particular work activities. Technological innovation
must first deliver the technical capability before a workplace solution can be developed
and deployed.
In Chapter 2, we outlined 18 performance capabilities required to carry out the range of
work activities and the current state of the technology for those capabilities as measured
against human performance. Along with our assessment of the current state, we developed
progression scenarios for each of these capabilities. We did so through surveys of academic
and industry experts and through an extrapolation of metrics including recent commercial
successes and the historical trajectory of the capabilities, along with a range of other
predictors. Details of our methodology can be found in the technical appendix.
While machines can already match median human performance or even exceed the
top levels of human performance in some of the 18 capabilities, such as, for example,
information retrieval, gross motor skills, and optimization and planning, many other
capabilities require more technological development, for example to raise natural language
understanding and logical reasoning to a median human level.
Scenarios for achieving higher levels of performance capabilities

Exhibit16 shows a potential range of time frames for technology to attain the next level of
performance for each capability. From a technical standpoint, some of these performance
capabilities are quite far advanced, and developing top-quartile human performance may be
relatively fast, although there are some significant variations among the different capabilities.
Sensory perception is already at median human performance. A robot tongue can
already detect the color index and alcohol content of beer with more than 80percent
accuracy, for example.69 For tactile perception, which can already exceed top-quartile
human performance for several dimensions, key challenges include miniaturizing the size
of hardware and adapting the sensors to function in different environments.70
Cognitive capabilities are considerably more varied in their performance compared
to humans. Information retrieval, optimization and planning, and recognizing known
patterns and categories have already reached the top quartile of human performance,
whereas generating novel patterns, creativity, coordination with multiple agents, and
logical reasoning and problem solving are still at a relatively early stage. In terms of
coordinating with multiple agents, robots already have demonstrated the ability to
coordinate with similar types of robots.71 Their ability to collaborate with humans is still at
an early stage. Automated creativity is perhaps the furthest away; computer creativity for
now requires human involvement to judge the quality of work or to provide direction.
Carlos A. Blanco et al., Beer discrimination using a portable electronic tongue based on screen-printed
electrodes, Journal of Food Engineering, volume 157, July 2015.
70
Giorgio Cannata and Marco Maggiali, Design of a tactile sensor for robot hands, in Sensors: Focus on
tactile, force and stress sensors, Jose Gerardo Rocha and Senentxu Lanceros-Mendez, eds., InTech, 2008.
71
Julia Sklar, Making robots talk to each other, MIT Technology Review, August 2015.
69
71
Social and emotional capabilities for now are below median human performance levels.
Despite advances in artificial intelligence, machines still have difficulty identifying social
and emotional states (sensing), drawing accurate conclusions about them (reasoning),
and responding with emotionally appropriate words or movements to them (output).
Physical capabilities are already at top-quartile human performance for gross motor
skills and navigation, which has enabled wide deployment of robots in industrial
automation, military, and defenseand given consumers turn-by-turn navigation apps
for their smartphones. For fine motor skills, we estimate that top-quartile performance
could be achieved when robotic hands have the same degrees of freedom as human
hands. Mobility remains a challenge, especially vertical mobility such as climbing stairs
and ladders.
Exhibit 16
Ranges of estimated time frames to reach the next level of performance for 18 human-related
performance capabilities
Rating
Below median
Median
Top quartile
Human performance
Median
Top quartile
Rating
Sensory
Sensory perception
Cognitive
capabilities
Recognizing known patterns/categories

Generating novel patterns/ categories
Creativity
Output articulation/ presentation
Natural
language
processing
Social and
emotional
capabilities

Physical
capabilities

Gross motor skills
Navigation
Mobility
2010
20
30
72
40
50
60
70
2080
Natural language capabilities are critical for automation across many activities
Natural language understanding is of critical importance for a wide range of applications in
the workplace. Despite some technical advances in this capability, including the accuracy
of machine translation, machines still have a long way to go to achieve median human
performance. The speed of technological advances in natural language capabilities in the
future will substantially affect the timelines for automation overall.
The development of higher levels of natural language understanding is the single most
important factor constraining the technical automation potential of activities for which
the current performance level of technologies is insufficient. In the earliest scenario for
automation, the current level of performance in natural language constrains the proportion
of all activities that can be automated by 18percent. In the latest scenario, the current level
of natural language understanding would hinder automation of 40percent of activities, since
other capabilities will have reached their required performance levels by then.
Natural language processing capabilities have advanced in recent years and will likely
continue to develop and improve, led by their use in auto (in-vehicle speech recognition),
health care (clinical documentation improvements), and general personal use (virtual
assistants in both mobile and fixed devices). Technology companies and venture capitalists
continue to invest heavily in technologies related to natural language processing. Algorithms
can write passages and new articles that are largely indistinguishable from those written by
humans, and many experts expect continued rapid technical advances; by 2018, according
to one prediction, machines could write as much as 20percent of all business content.72
HISTORICAL TIME FRAMES FOR DEVELOPING SOLUTIONS THAT INTEGRATE

MULTIPLE CAPABILITIES
Many of the 18 performance capabilities in our framework operate together. To arrange
tables or dining areas as well as a human waiter, for example, machines will need some
capacity for natural language understanding and generation, an ability to recognize known
patterns, retrieve information, coordinate with multiple agents, and navigate. Furthermore,
machines will need optimization and planning capabilities, fine motor skills, gross motor
skills, and mobility. Developing solutions for this type of multi-capability activity requires time
and technology.
To develop scenarios for the times required for solution development, we looked at historical
precedents, examining almost 100 existing automation solutions that use hardware or
software or both. We collected details of the development time and the technical capabilities
that were integrated, recording the number of years from the initial research to the product
launch, and identifying as many as three of the most relevant capabilities out of our 18
that were used. For example, Viv, an artificial intelligence technology that aspires to be
the intelligent interface for everything, combines logical problem solving with natural
language understanding and generation, and took four years to integrate.73 PillPick, a
pharmacy automation system that helps hospitals eliminate the risk for medication errors
during packaging and dispensing, combines fine motor skills with sensory perception, and
took two years to develop.74 For each activity for which a solution needs to be developed,
we used the 25th and 75th percentile time frames for the capabilities whose historical
development times at those points were longest. For more details of our methodology, see
the technical appendix.
Bernard Marr, Why management dashboards and analytics will never be the same again, Forbes, January
21, 2016; Gartner predicts our digital future, Gartner, October 6, 2015.
73
Lucas Matney, Siri-creator shows off first public demo of Viv, the intelligent interface for everything,
TechCrunch, May 9, 2016.
74
Anthony Vecchione, New automated packaging system ensures patient safety, Drug Topics, November
22, 2004.
72
73
On average, we found that solution development takes one year for our earliest scenario
and nine years for the latest. Some of the longest time frames were for social and emotional
capabilities, whereas a number of the cognitive capabilities can be integrated much more
rapidly (Exhibit17).
Exhibit 17
Development timelines for solutions associated with each capability
Minimum
Maximum
25th percentile
50th percentile
75th percentile
Range of solution development time frames (years)
Sensory
Sensory perception
Cognitive
capabilities
Recognizing known patterns/categories

Creativity
Output articulation/presentation
Natural
language
processing
Social and
emotional
capabilities

Physical
capabilities

Gross motor skills
Navigation
Mobility
0
10
MODELING THE ECONOMIC FEASIBILITY OF AUTOMATION

Our model assumes that the adoption of automation will begin after it becomes
economically feasiblethat is, when the economic benefits exceed the costs. These
benefits include not only labor cost reductions, but also other performance improvements
such as higher throughput or improved quality, as discussed earlier in this chapter. However,
particularly early in the adoption cycle, the certainty that these benefits can be captured
will be lower until they are demonstrated at scale. Thus, for the purposes of modeling, we
assume that adoption begins when the cost of automating a particular activity (for a specific
occupation in a country) is at parity for the cost of human labor for the same activity. To be
fair, the certainty that labor cost savings can be completely captured will also be somewhat
lower early in the adoption cycle. On a net basis, our model assumes that decision74
makers discount the benefits of initial labor cost savings by roughly the same amount as
they believe the also uncertain non-labor cost-related benefits will be captured. In our
experience, business leaders tend to view labor cost savings as being more predictable
than other benefits.
Moreover, as we will discuss in the section on the international impact of automation, the
overall lower level of wages in some emerging economies could mean that automation
proceeds more slowly there than in advanced economies, where wages are higher than the
global average. The cost of labor is not necessarily fixed, however. The related labor supply
and demand dynamics are also a factor in the timeline for automation, as discussed above,
but we do not model these dynamics.
ADOPTION OF AUTOMATION TECHNOLOGIES AFTER COMMERCIAL

AVAILABILITY CAN TAKE YEARS
Even when a solution to automate an activity is commercially available, and implementing
the solution makes rational economic sense, adoption still takes timeand can be very
costly. Full adoption across an entire sector of any technology, particularly those that are
integrated into the workplace, takes years. Individual decision makers must become aware
of the potential, and there is always a spectrum of willingness to adopt, a phenomenon
well documented in the research studying the diffusion of innovations.75 Capital has to be
deployed, technology acquired and installed, and processes transformed, within or across
enterprises. Often, regulations need to change, and individual workers and employees have
to become accustomed to the new technologies and processes.
From a review of the historical rate of adoption of previous technologies, the time from
commercial availability to 90percent adoption ranges from approximately eight to 28years
(for 50percent adoption, the range is between about five and 16years). This lag applies
not only to hardware-based technologies that are capital-intensive and require physical
installation; even technologies that are made available purely online take years to be
adopted. A fast-growing consumer service such as Facebook began in 2004, and as of
this writing, has not yet reached full adoption, even of non-China internet users (see Box4,
Adoption of hardware vs. software/cloud-based technologies). Furthermore, that decade
of adoption does not even take into account previous instances of social networking
services. Automation technologies that would be incorporated into the workplace require
even more user and process changes than a consumer service that individuals can
adopt independently.
To create scenarios for the rate of automation adoption, we analyzed adoption rates of 25
previous technologies to establish a range of timelines. We incorporated these historical
examples into classic S-shaped adoption curves. (Exhibit18).76 We analyzed both
hardware and software/online technologies, and we divided them into groups representing
technologies with the fastest and the slowest adoption. Technologies with the fastest
adoption rates include stents, airbags, MRIs, TVs, and online air booking. The slower
adoption category include cellphones, personal computers, dishwashers, and pacemakers.
Everett M. Rogers, Diffusion of innovations, The Free Press, 1962.

S-shaped adoption curves were studied for consumer technologies by Frank Bass, Comments on A
new product growth model for consumer durables, Management Science, volume 50, number 12,
December 2004.
75
76
75
Box4. Adoption of hardware vs. software/cloud-based technologies

Hardware-based automation technologies, such as robots and self-driving
vehicles, have requirements that can lengthen the time of adoption, that is,
substantial capital requirements and the need to physically produce and
deploy these technologies.
These requirements are much lower for software-based technologies,
particularly those that are deployed through cloud technologies, that is,
where the bulk of the computing occurs at centralized data centers that are
accessed through networks. For these technologies, customers often pay for
these services on an on-demand basis, reducing the need for the customer
to deploy capital expenditure or manage for peaks in capacity (the cloud
provider takes on these tasks). In general, the marginal cost of producing one
more instance of a piece of software tends toward zero. And there are several
cases of specific pieces of consumer applications, particularly those that have
gone viral, whose adoption seems to have been extraordinarily fast, such as
Pokemon Go.
Does this mean that we should expect the adoption of software/cloud-based
automation technologies to be much faster than for other technologies?
In examining the historical record, we find that the adoption of softwarebased technologies falls within the envelope of the adoption rates of other
technologies, that is, ranging from eight to 28years to reach close to full
adoption. In particular, for technologies that will be implemented in the
workplace, the technical deployment of the technology represents only a
fraction of the time and cost necessary to embed the technology into the
processes and practices of an enterprise. We examined the adoption of
cloud-based versions of enterprise resource planning (ERP), supply chain
management (SCM), and customer relationship management (CRM). While
these technologies have not yet reached a plateau in adoption, their adoption
curves all fall within the range of adoption of other technologies. For instance,
cloud CRM was first offered by Salesforce in 1999, and adoption of that
technology still continues to grow.
Even for consumer online technologies, for which the bureaucracy of an
enterprise is not a barrier to adoption, the behavior changes necessary for
adoption take time to proliferate, particularly when these technologies are
appropriately viewed by category, rather than a specific service (for example,
Pokemon Go was not the first mobile game, and has not yet come close to full
adoption). Take consumer social networking as another example. Facebook,
only one example of a social networking platform, and not the first, was
launched in 2004, and over a decade later, adoption continues to grow, both of
that platform and other social networking platforms. Peer-to-peer (P2P) mobile
payments is another consumer technology whose adoption rate falls within the
envelope of other technologies.
76
Exhibit 18
Historic adoption curves for technological innovations
Adoption trend by technology1
Fast case
Slow case
% of full adoption potential

100
90%
adoption
90
80
70
60
50%
adoption
50
40
30
20
10
0
10
15
20
25
30
35
40
45
Years since adoption
1 Technologies considered include airbags, antilock braking systems, cellphones, cloud CRM, cloud ERP, cloud SCM, color TVs, copper production through
leaching, dishwashers, electronic stability control, embolic coils, Facebook, instrument landing systems, laparoscopic surgery, Lithium-ion cell batteries,
microwaves, MRI, online air booking, P2P remote mobile payment, pacemakers, PCs, smartphones, stents, TVs, and VCRs.
The adoption of technologies within enterprises include factors beyond those that underpin
consumer adoption. For example, the technology adoption literature discusses rank effects
(that is, how the different individual characteristics of firms, such as their size, can affect
the rate and extent to which they adopt new technologies) and the effects of competitive
dynamics (that is, how the adoption of new technology by one company in an industry
could influence the adoption of technology by other companies in that industry).77 We do
not model firm-level adoption dynamics; our model takes advantage of the fact that at a
high level for economies and industries, the net result of the enterprise adoption factors are
S-shaped curves that resemble those for consumer adoption.
This combination of technical feasibility, solution development, economic feasibility, and
adoption enables us to model a set of scenarios that encompass various time frames for
the pace of automation. To illustrate, we detail a specific example, driving heavy trucks (see
Box5, Driving heavy trucks: Modeling scenarios for adoption of automation).
Adoption will ultimately depend on several factors include overcoming possible policy
barriers and resistance to automation, which are meant to be captured in scenarios in the
adoption stage of the model. For instance, given the importance of technological advances,
intellectual property or regulatory issues could delay certain companies from being able to
deploy specific technologies.
See Massoud Karshenas and Paul Stoneman, Rank, stock, order, and epidemic effects in the diffusion
of new process technologies: An empirical model, RAND Journal of Economics, volume 24, issue 4,
1993, and Marc Anthony Fusaro, The rank, stock, order and epidemic effects of technology adoption: An
empirical study of bounce protection programs, The Journal of Technology Transfer, volume 34, issue 1,
February 2009.
77
77
Box5. Driving heavy trucks: Modeling scenarios for adoption

of automation
To illustrate how we model scenarios for the pace and extent of automation,
let us consider the single but widely performed work activity of driving, for one
common occupation: heavy and tractor-trailer truck drivers.1 This occupation
is a significant source of employment, with more than 20million workers
worldwide, of whom 3.2million are in India, 2.9million are in China, 1.6million
are in the United States, and one million are in Japan.2
From a purely technical standpoint, the technologies to automate the
required capabilities for this activity already exist. Some of the 18 individual
performance capabilities we use in our framework are not required for this
activity, including the three social and emotional capabilities, but for other
capabilities, existing technology has demonstrated the necessary constituent
performance levels to potentially enable driving. However, in order to automate
the activity of driving (Level 4 autonomy as classified by the Society of
Automotive Engineers), these individual capabilities must be integrated into a
solution, in other words the hardware and software that would constitute an
autonomous driving truck must still be engineered.3 Based on the historical
time frames required to create solutions that involve the capabilities necessary
to automate this activity, we estimate the time required to engineer such a
solution could take more than seven years, which we use as basis for our latest
adoption scenario. Conversely, because existing automation technology has
met the levels of performance required across all of the necessary capabilities,
and we do not know how far along pre-existing engineering efforts might be,
our model also accepts the possibility that an organization could announce
a Level 4 autonomous truck immediately, which we use as the basis for our
earliest adoption scenario.
Then comes the question of economic feasibility. In our model, adoption
begins when the cost of automation reaches parity with the cost of human
labor, for an individual occupation within a specific market (which we model
at the level of countries). Of course, there are other economic benefits which
could contribute to the business case for automation. In the case of driving
heavy trucks, for example, these could include higher fuel efficiency, improved
safety, increased asset utilization, and so on. However, particularly early in the
adoption cycle, the certainty that these benefits can be captured will be lower
until they are demonstrated at scale. To be fair, the certainty that labor cost
savings can be completely captured will also be somewhat lower early in the
adoption cycle. On a net basis, our model essentially assumes that decisionmakers discount the benefits of initial labor cost savings, because of initial
uncertainty, by roughly the same amount as they believe the also uncertain
non-labor cost-related benefits will be captured.
The detailed work activity title is operating vehicles or material-moving equipment in the US
Bureau of Labor Statistics taxonomy.
2
US Bureau of Labor Statistics; Oxford Economic Forecasts.
3
Taxonomy and definitions for terms related to driving automation systems for on-road motor
vehicles, Society of Automotive Engineers, September 30, 2016.
1
78
Box5. Driving heavy trucks: Modeling scenarios for adoption

of automation (continued)
Using these assumptions, the time at which we model adoption beginning
is sensitive to the cost of labor in different markets. For our example, we
compare the United States with China (Exhibit19). Total wages paid to heavy
and tractor-trailer truck drivers exceed $328billion worldwide, of which almost
$60billion is paid in the United States, and more than $33billion is paid in
China. Because of the significantly higher level of wages in the United States,
automation is modeled to become economically feasible faster there. Under
our scenarios, the cost of the automation could be below US wage levels
within three to ten years after Level 4 autonomy is available, whereas in China it
could take 10 to 16years.
Finally, once automation becomes economically feasible, adoption can still
be a long process, even when the business case is compelling. In this case,
there are approximately two million tractor-trailers in the United States, each
of which is typically on the road for about 20years, and cost about $160,000
apiece. Replacing the current fleet would cost $320billion, not including the
additional technology for autonomous driving (whose cost we do model as
declining over time), and would be certain to take many years.
Exhibit 19
Modeled adoption of operating vehicles or material-moving equipment, United States and China1
% of adoption vs. year
United States earliest
United States latest
China earliest
China latest
100
90
80
70
60
50
40
30
20
10
0
2015
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95 2100
05
2110
1 The detailed work activity (DWA) is 4.A.3.a.4.I01.D06. Occupations using this activity include: Heavy and Tractor-Trailer Truck Drivers, Industrial Truck and
Tractor Operators, etc.
There could be a potential retrofit opportunity for more recent vehicles with electronic control systems, and
perhaps the number of autonomous trucks necessary to meet current demands could be lower than the
current number of trucks because of higher utilization. However, the basic point remains, that adoption will
take years and large amounts of capital to reach its maximum level.
Box 5
79
MODELING THE PACE AND EXTENT OF AUTOMATION ACROSS ACTIVITIES

AND SECTORS
Each of the four stages affects the overall pace of automation reflected in the model.
Technical feasibility accounts for much of the variance in our modeled scenarios, but
economic feasibility is also a significant factor, especially in the earliest scenario, where our
modeling suggests it could hold up adoption and deployment for eight to nine years for
some activities (Exhibit20).
Several of the 18 performance capabilities are potential bottlenecks in the model. The
types of activities that today have among the lowest potential for automation, such as
managing and developing people, applying expertise and experience, and interfacing
with stakeholders, will all require substantial further advances in social and emotional
capabilities, as well as in natural language understanding and creativity. These activities
typically require at least a median-level human performance in social and emotional
capabilities, for example. As we have seen, these capabilities are particularly difficult
to develop, and most related technologies are in nascent stages and may take several
decades to mature. This potentially could delay the development, integration, and
deployment of automation as a whole.
Along with timelines for capability development, whether hardware or just software is
required is also a key determinant of scenarios in the model, because this will affect the
economic viability of automation. Hardware requires significant capital spending, and thus
we model with relatively higher initial costs. For example, sensory perception capabilities
need cameras and sensors. Mobility requires wheels or other hardware that enable
machines to move. Software solutions, by comparison, are relatively less expensive to
deploy and thus we model with relatively lower costs compared to hardware solutions.
Some 98percent of the solutions for predictable physical activities are hardware ones,
whereas hardware represents only about 30percent of processing data solutions.
This hardware-software distinction helps explain differences in the pace of automation
adoption across different types of activities. For example, under our earliest scenario, the
model suggests that unpredictable physical labor activitieswhich require substantial
hardware for successful automationwill have much slower adoption rates than for
activities that consist of processing data, which primarily require software solutions.
As we will detail in our discussion of country differences below, the modeled time frame
for physical activities is longer in emerging economies largely because of lower wages
compared to the cost of hardware-based automation solutions.
80
Exhibit 20
Differences in the adoption timing are primarily driven by technical feasibility, and in the earliest scenario,
also by economic feasibility
Decomposition of average time to 50% adoption into stages by types of activities, global
Years
Adoption
Earliest
scenario
27
5
5
25
24
2
3
13
69
22
3
3
15
14
Latest
scenario
17
5
1
16
14
5
2
49
48
67
62
16
16
54
16
11
16
11
43
16
16
16
13
13
12
7
33
13
6
32
30
6
7
18
14
14
7
Share of time
automatable1
%
Managing and Applying

developing
expertise
people
Interfacing
with
stakeholders
Unpredictable Collecting
physical
data
labor
Processing
data
Predictable
physical
labor
14
24
38
73
74
23
72
1 Percentage of time automatable by adapting currently demonstrated technology.

81
GLOBAL TIMELINES: MODELING ADOPTION OF AUTOMATION BY COUNTRY

Automation will be a global force that affects all countries, whether they are emerging
economies or advanced ones. In the earliest scenario we modeled, automation could
account for more than 50percent of working hours in two-thirds of countries within just
20years, by 2036. In the latest scenario we modeled, more than half of all countries will have
50percent automation or more within 50years, by 2066 (Exhibit21).
In the early years of automation adoption, our model shows the impact being felt most
strongly in a few advanced economies, especially Germany, Japan, and the United States.
These countries have both high wages and major industries that already have a high
potential for automation based on existing technologies. As automation is adopted in more
countries around the world, the impact will become especially pronounced in China and
India, because of their large workforce. Initially, our model shows automation affecting
workers there in manufacturing and retailing because of their high automation potential, but
in the longer term its largest impact will be in agriculture, which is where hundreds of millions
of Chinese and Indians earn their livelihood.
Exhibit 21
Automation impact will be global under multiple modeled scenarios
% of work hours automated
100
2036
2066
51%
99%
2%
46%
Early
scenario
Latest
scenario
SOURCE: US Bureau of Labor Statistics, 2014 O*Net database; McKinsey Global Institute analysis
82
No data
Adoption of automation could be faster initially in advanced economies than emerging

ones because of wage levels and integration solution costs. Higher wages in advanced
economies and hardware costs will likely make automation economically viable faster
there than in emerging economies. In Japan, the United States, and Europes five largest
economiesFrance, Germany, Italy, Spain, and the United Kingdomthis is likely to mean
earlier adoption in a range of sectors, especially manufacturing and services. In the United
States, for example, our model shows manufacturing, retail, health care, transportation,
accommodation and food services, and administrative services to be among the first
sectors affected. In terms of the impact on workers, sectors that are large employers,
such as health care, will be affected even though their expected automation adoption rate
(48percent) is lower than some other high-adoption sectors such as accommodation and
food services (83percent) in 2036 under earliest scenario.
Initially, technologies with physical capabilities are the most likely to become available, and
will be increasingly adopted in sectors such as manufacturing and retail. As adoption in
these sectors approaches 100percent, their contribution to overall adoption rates will hit
a plateau. Once the technology to replicate cognitive and natural language understanding
capabilities has been mastered, services sectors globally will be affected, and the pace
and extent of overall automation will pick up, especially in advanced economies with large
service sectors.
For all advanced economies, the business case for automating will become stronger as
solutions for technology integration in the workplace become cheaper. Exhibit22 shows the
estimated wages corresponding to automation adoption by country.
In China, India, and other emerging economies, cost and relative lower wage levels will
likely delay adoption. Manufacturing relies heavily on predictable physical activities, and
automating them will require hardware solutions that require considerable up-front capital
investment. This may not be economical in emerging economies, given the lower cost of
labor there, until the cost of the solutions drops sharply.
Emerging economies could achieve a pace of automation similar to that of advanced
economies if solutions become cheaper, possibly through localized innovation. Adoption
could also be accelerated as a result of policy measures, increased competition, a lack of
legacy systems that could be a brake on automation implementation, and a high degree of
technological literacy.
In sectors where software solutions to integrate automation technologies will be required,
the pace of automation across emerging and advanced economies could be similar. For
example, we model similar rates of automation adoption in the finance and insurance
sectorswhich are characterized by a high proportion of data processing and collectionin
both the United States and China. Software, which has a relatively minimal marginal cost,
accounts for just over half the technology integration solutions needed in this sector, and
the global disparity in wages is not as pronounced in some areas. For example, architects
or investment bankers in emerging economies are relatively well paid. Moreover, wage
distribution globally is fairly similar for real estate, rental and leasing activities, health care,
and social assistance.
83
Exhibit 22
Wages associated with automated activities are modeled to be highly concentrated in countries with higher wages
and populations
Wages associated with automated activities modeled1
$ trillion
Top 4
countries
Early scenario
5.5
(18%)
2036
6.6
(22%)
Total =
30.3
2066
27.8
(22%)
Total =
125.8
Remaining
countries
Late scenario
1.7
Japan
0.1
Germany
1.9
India
0.2
Japan
5.3
United
States
0.2
China
0.9
United
States
0.3
(13%)
0.6
(27%)
18.2
(60%)
21.6
(17%)
Other
top 10
Total =
2.2
1.3
(61%)
9.2
China
6.0
Japan
3.5
Japan
12.0
India
5.0
India
16.1
United
States
8.2
United
States
20.4
China
10.0
(17%)
13.3
(22%)
76.4
(61%)
42.3
Total =
60.3
China
37.0
(61%)
1 Forty-six countries used in this calculation, representing 78% of global labor force; calculations based on current FTEs in each sector, not considering FTE
growth and sector migration.
NOTE: Not to scale. Numbers may not sum due to rounding.
84
The largest impact on workers could be felt in labor-intensive sectors in China

and India, whereas in other countries it will be felt across multiple sectors
In analyzing the potential impact of automation on workers, we make a number of
assumptions. We calculate the impact based on the current number of full-time equivalents
in each sector and do not take into account any possible growth in their numbers per sector
or the movement of labor from one sector to another. We also assume that the technologies
we identify will emerge throughout the world and that they will be adopted only when they
become economically feasible.
China and India both have very large farming sectors, with about 230million people out
of a working population of 450million in India alone working in agriculture. In China, more
than 200million work on the land out of a total workforce exceeding 770million. Given the
employment size of this sector, even a relatively low rate of automation adoption of about
10percent could have significant employment consequences in both countries.
In both China and India, the impact of automation on employment could also be felt in the
retail and manufacturing sectors, as both have a relatively high potential for automation and
a sizable labor force. In the five major European countries, Japan, and the United States, the
employment impact will likely be spread across multiple sectors, especially in the event that
large-scale automation begins relatively soon. A detailed view of our model of automation
impact on individual countries is available via an interactive graphic online.78
Automation technologies will be increasingly adopted in every industry, every sector,

and every country in the world, but there is considerable uncertainty about the speed
and intensity with which they will arrive. These will depend on several key variables, both
technical and economic. Machines will need to be able to simulate the full range of human
performance capabilities, and solutions to integrate the technology into the workplace will
need to be adopted. Only when costs have fallen below wage levels will automation become
economically viable, and ultimately adopted. All of this could happen within two decades
for a wide range of activities and sectors, but it could also take much longer. Whatever the
time frame, the consequences will be significant not just for individual workers and sectors
but also for the global economy as a whole. In the next chapter we look at how automation
could upend some cherished notions about productivity, growth, and development.
The data visualization can be viewed online at http://public.tableau.com/profile/mckinsey.analytics#!/
78
85
Laboratory/Alamy Stock Photo

86
5. AN ENGINE OF PRODUCTIVITY
The world is in need of a new engine of GDP growth. Shifting demographics in many
countries brought on by aging populations and declining birthrates are reducing the share
of working-age populations and creating an economic growth gap: in the not-so-distant
future, without an acceleration in productivity growth, there will not be enough workers for
countries to meet their aspirations for growth in GDP per capita.
In this context, automation could be a significant opportunity. Our estimates suggest it could
help serve as a new productivity engine for the global economy, bridging that economic
growth gap. Automation could raise global productivity by as much as 0.8 to 1.4percent
annually. While that will not be enough to ensure all countries meet their per capita GDP
growth aspirations, it will make a major contribution toward that goal.
In order for this growth to take place, people will need to keep workingalongside the
robots that will help provide the productivity boost. Even with automation, a deficit of labor
is more likely than a surplus. Yet the adoption of automation will change the nature of work,
and the public debate about it often focuses on the prospect that it could lead to very
large-scale unemployment. Such fears are not new: already back in 1966, a US government
commission noted concerns that technological change would in the future not only cause
increasing unemployment, but that eventually it would eliminate all but a few jobs, with the
major portion of what we now call work being performed automatically by machine.79
In fact, the large-scale shifts in employment that automation will enable are of a similar
order of magnitude to the long-term technology-enabled shift in the developed countries
workforces away from agriculture in the 20th century. That movement did not result in
long-term mass unemployment because it was accompanied by the creation of new
types of work not foreseen at the time. We cannot definitively say whether things will be
different this time. But our analysis does show that automation will fundamentally alter the
workplace, requiring all workers to cohabit extensively with technology and reshaping the
corporate landscape.
THE PRODUCTIVITY BOOST FROM AUTOMATION COULD BRIDGE A LOOMING

ECONOMIC GROWTH GAP
GDP growth was exceptionally brisk over the past half century, driven by the twin engines
of employment growth and rising productivity. However, declining birthrates and the
trend toward aging in many advanced and some emerging economies mean that peak
employment will occur in most countries within 50years. The workforce in Japan is already
shrinking in size, and the total number of workers in China will start to decline within a
decade. This expected decline in the share of working-age population will place the onus
for future economic growth far more heavily on productivity gains. Employment growth of
1.7percent annually between 1964 and 2014 in the G19 countries and Nigeria is set to fall to
just 0.3percent per year.80
Technology and the American economy, Report of the National Commission on Technology, Automation and
Economic Progress, US Department of Health, Education and Welfare, February 1966.
80
The global productivity challenge created by waning employment growth is detailed in our report Global
growth: Can productivity save the day in an aging world? McKinsey Global Institute, January 2015. The
G19 countries are the G20 minus the European Union: Argentina, Australia, Brazil, Canada, China, France,
Germany, India, Indonesia, Italy, Japan, Mexico, Russia, Saudi Arabia, South Africa, South Korea, Turkey, the
United Kingdom, and the United States. Including Nigeria, these 20 countries generate more than 80percent
of global GDP.
79
Prior MGI research has shown that even if productivity growth maintains its 1.8percent
annual rate of the past half century, the rate of GDP growth will fall by as much as 40percent
over the next 50years. On a per capita basis, the GDP growth decline is about 19percent
(Exhibit23). In order to compensate for slower employment growth, productivity would need
to grow at a rate of 3.3percent annually, or 80percent faster than it has grown over the past
half century.81
Exhibit 23
Without an acceleration in productivity growth, demographic trends will cut GDP growth by nearly half,
causing a decline in historic GDP per capita growth rate
Historical vs. future GDP and GDP per capita growth
Compound annual growth rate, 19642014 historic and 201565 projected
%
GDP growth
Productivity growth
Employment growth
3.5
-46%
1.8
2.1
1.9
1.6
1.8
1.8
1.7
1.8
0.3
0.1
Past 50 years
(19642014)
-19%
-0.2
Next 50 years
at historical
productivity growth
(201565)
Past 50 years
(19642014)
Next 50 years
at historical
productivity growth
(201565)

The size of the workforce over the next 50years is too small to maintain current
per capita GDP growth without accelerating productivity growth
An economic growth gap has opened up as a result of the shrinking workforce. The number
of full-time equivalents needed just to maintain the current GDP per capita over the next
50years is larger than the number of workers who will be available, given demographic
trends in most countries. We estimate this gap in economic output as being about
130million full-time equivalents (FTEs) for the G19 countries and Nigeria alone. (In other
words, the economic output equivalent to an additional 130million full-time workers would
be needed to maintain current GDP per capita, assuming no productivity gains.) If countries
are to achieve more ambitious longer-term growth in line in line with their development and
81
88
Our estimate of employment growths contribution to GDP growth in this report differs slightly from this earlier
research, as we have assumed productivity measured in each country, rather than a global average.
5. An engine of productivity
the aspirations of their citizens goals (that is, growth in GDP per capita), the gap would be
considerably largerthe economic output of about 6.7billion FTEs by 2065 (Exhibit24).
Exhibit 24
Demographic trends are creating a pressing need for an acceleration in productivity growth
%
Historical
19642014
Future
201565
3.5
2.9
Productivity
growth
1.8
2.8
Employment
growth
1.7
0.2
0.1

%
Billion
0.1
0.1
Historical
Required to
maintain current
GDP per capita
Required to
achieve projected
GDP per capita
2.1
2.5
0.1
6.7

For the purposes of our analysis, we based our country-level GDP projections on
McKinseys proprietary Global Growth Model. This model projects a global GDP growth
rate of 2.9percent, resulting in an annual productivity growth of 2.8percent. For advanced
economies overall, the projections in the McKinsey model result in a compound annual
growth of GDP per capita to 2030 of 1.4percent, and 3.4percent for emerging economies.
SIMILAR from ES
(no RH cols with orange)
89
For the G19 countries and Nigeria that we discuss in this report, the model projects GDP
growth of 2.7percent, resulting in GDP per capita growth of 2percent to 2030.82
The growth gap between the projected growth and growth that must be provided by
productivity increases is most pronounced in fast-growing countries such as China, India,
Indonesia, and Nigeria, but it is also prevalent in Germany and Japan and other countries
that are already experiencing a slowdown or decline in the working-age population.
Furthermore, productivity growth has slowed in many countries.83
Exhibit25 illustrates the need for continued productivity growth, even just to maintain GDP
per capita (an admittedly unsatisfactory outcome) in both the medium term (by 2030) and
in the long term (by 2065) for the 20 largest economies in the world. Without productivity
growth, aging countries with older demographics simply would not have enough workers
needed to maintain GDP per capita. Younger countries have more than enough workers to
maintain GDP per capita.
Automation can help bridge the projected growth gap by compensating for the
slowdown in workforce growth
Our analysis of the automation adoption scenarios suggests that automation could help
bridge the projected growth gap caused by a deficit of full-time equivalents worldwide.
Automation alone will not be sufficient to achieve long-term target growth across the world,
given the decline in the working-age population and the need for high productivity to achieve
that target. Especially in fast-growing countries, other measures to boost productivity will be
needed. However, notably, the productivity gains from automation could suffice to at least
maintain todays GDP per capita.
Our methodology takes into account only labor substitution gains. Other performance
gainsin the form of improved quality, fewer breakdowns, greater safety, and so onwould
come on top of this overall productivity boost. We also assume that human labor displaced
by automation would rejoin the workforce and be as productive as it was in 2014, that is,
new demand for labor will be created. In some ways, this is a conservative assumption,
given that if automation produces productivity gains, we assume displaced labor reenters
the workforce at a lower level of productivity than the level of labor productivity at the time
the displacement occurs. Others could argue that the activities performed by workers who
are displaced by automation could have lower levels of economic output than the activities
that had been taken over by machines. In any case, it is vital that there be new demand for
labor displaced by automation.
McKinsey & Companys proprietary Global Growth Model provides complete time-series data for more
than 150 concepts and 110 countries over 30years. It incorporates more than a dozen major international
databases from such institutions as the United Nations, the World Bank, the International Monetary Fund,
and the Bank for International Settlements. The structure of the model emphasizes the drivers of economic
growth, including demographic factors, education, energy supply, physical capital, and some determinants
of total factor productivity. It captures the long-term effects of urbanization and industrialization, as well as
the impact of sociopolitical institutions, especially on finance and governance. Because business-cycle
fluctuations affect growth in the short term, the model also links trade and international capital flows, credit
and asset markets, and the monetary relationships that determine inflation, interest rates, and exchange rates.
See the technical appendix for more details. The growth model shows a compound annual growth rate for
GDP and GDP per capita for 2015 to 2030 by country as follows: China 5.1percent GDP, 4.9percent GDP
per capita; France 1.0percent and 0.6percent; Germany 1.3percent and 1.6percent; India 5.8percent and
4.8percent; Italy 0.5percent for both GDP and GDP per capita; Japan 1.0percent and 1.3percent, United
Kingdom 1.5percent and 1.0percent; United States 2.2percent and 1.4percent.
83
82
90
Exhibit 25
Aging countries have the greatest need for productivity growth just to maintain GDP per capita
FTE gap between FTE projections and number of FTEs to maintain current GDP per capita, 2030 and 2065
Gap indexed as % of number of FTEs in 2014
FTE deficit
Medium term, 2030
France
-0.16
-0.12
-0.07
Germany
Italy
-0.14
-0.09
-0.15
-0.07
Japan
-0.09
-0.03
South Korea
-0.12
-0.02
United Kingdom
-0.04
United States
-0.04
-0.10
-0.06
-0.04
0.04
Brazil1
-0.11
-0.02
China
-0.08
-0.05
Russia
Indonesia
Mexico
0.16
0.10
0.11
0.11
0.06
Nigeria
South Africa
0.84
0.19
Saudi Arabia
Turkey1
0.23
0.16
India
Young
emerging
economies
0.03
0.06
Argentina1
Aging
emerging
economies
-0.13
-0.10
Canada
Aging
developed
economies
Long term, 2065
-0.06
Australia
FTE surplus
-0.02
0.19
0.10
0.05
-0.02
1 Countries on borderline of demographic categorization between aging and growing populations (under 1 million FTE surplus/deficit in 2065) categorized
according to demographic statistics and grouping of similar regional economies.
SOURCE: The Conference Board Total Economy database; International Labour Organisation; United Nations Population Division; Statista; McKinsey Global
Institute analysis
91
Only considering the labor substitution effects on current work activities, and assuming
that displaced labor reenters the workforce at 2014 levels of productivity, we estimate
that, by 2065, the productivity enabled by automation could potentially increase economic
growth by 0.8percent to 1.4percent annually, the equivalent of 1.1billion to 2.3billion FTEs.
(Exhibit26).
Exhibit 26
Globally, automation could become a significant economic growth engine as employment growth wanes
%
Historical
19642014
Future
201565
3.5
2.9
Productivity
growth
1.8
2.8
Employment
growth
1.7
1.4
0.2
0.1

%
1.5
0.8
0.1
0.1
0.1
0.1
Historical
Required to
maintain current
GDP per capita
Required to
Early scenario
Late scenario
achieve projected
GDP per capita
2.1
2.5
0.1
6.7

Billion
0.51.1

REPEATS in report
92
0.9
Exhibit27 shows what the effect of automation could be on the gap between growth targets
and economic output by country for the G19 and Nigeria. We discuss individual countries
and groupings of countries in more detail later in this chapter.
Exhibit 27
Modeled effects of automation on the gap between growth targets and economic output by 2030
Normalized as % of 2014 GDP
Economic output deficit assuming fixed 2014 productivity

Economic output surplus/deficit assuming earliest automation adoption
Economic output deficit assuming latest automation adoption
-130 -120 -110 -100 -90 -80 -70 -60 -50 -40 -30 -20 -10
10
20
30
40
Australia
Canada
France
Aging
developed
economies
Germany
Italy
Japan
South Korea
United Kingdom
United States
Argentina
Aging
emerging
economies
Brazil
China
Russia
India
Indonesia
Mexico
Young
emerging
economies
Nigeria
Saudi Arabia
South Africa
Turkey
NOTE: Surplus/deficits calculated as gap to projected GDP in 2030 as a percentage of 2014 GDP by using population projections; base case (no automation)
40
assumes 2014 productivity. Assumes no-130
other productivity growth than that provided by automation.
SOURCE: The Conference Board Total Economy database; International Labour Organization; United Nations Population Division; McKinsey Global Institute
analysis
93
While our estimates of the productivity boost from automation are substantial, they are of an
order of magnitude comparable to major technologies that have been introduced in the past
two centuries. For example, between 1850 and 1910, the steam engine has been estimated
to have enabled productivity growth of 0.3percent per annum. Analyses of the introduction
of robots in manufacturing and IT estimate that they have accounted for annual productivity
increases of 0.4percent and 0.6percent, respectively (Exhibit28).84 One difference is
that automation of current work activities as we have analyzed it encompasses multiple
technologies, not just one.
Exhibit 28
Automation of existing activities could increase productivity at magnitudes similar to other major technologies
Productivity growth
%
Earliest scenario
Latest scenario
2.2
1.4
0.6
0.8
0.3
0.4
0.3
Automation
(201530)
Automation
(201565)
Steam engine
(18501910)
Robots
(19932007)
IT
(19952005)
NOTE: We include multiple technologies in our analysis of "automation," so these technologies are not entirely comparable, but meant to provide an order of
magnitude comparison.
SOURCE: Nicholas Crafts, Steam as a general purpose technology: A growth accounting perspective, Economic Journal, volume 114, issue 495, April 2004;
Mary OMahony and Marcel P. Timmer, Output, input, and productivity measures at the industry level: The EU KLEMS database, Economic
Journal, volume 119, issue 538, June 2009; Georg Graetz and Guy Michaels, Robots at work, Centre for Economic Performance discussion paper
1335, March 2015; McKinsey Global Institute analysis
The productivity potential of automation is a multiplier of other productivity levers.

Economies that have relatively low productivity can also accelerate productivity growth
through other means, for example by adopting best practices from other countries with
high productivity, in order to reap the full benefits of automations potential and achieve a
faster growth trajectory. Previous MGI research has estimated that about three-quarters
of potential productivity improvements could come from the broader adoption of best
practices and technologies, as companies catch up with sector leaders. The remaining
one-quarter would come from technological, operational, and business innovations that go
beyond best practices and push the frontier of the worlds GDP potential. Business leaders
and policy makers can encourage this acceleration of productivity by removing barriers
to competition in service sectors, investing in physical and digital infrastructure, exploiting
Nicholas Crafts, Steam as a general purpose technology: A growth accounting perspective, Economic
Journal, volume 114, issue 495, April 2004; Mary OMahony and Marcel P. Timmer, Output, input, and
productivity measures at the industry level: The EU KLEMS Database, Economic Journal, volume 119,
issue 538, June 2009; George Graetz and Guy Michaels, Robots at work, Centre for Economic Performance
discussion paper number 1335, March 2015.
84
94
data to identify transformational improvement opportunities, opening economies to crossborder flows, and crafting a regulatory environment that fosters increased productivity and
innovation.85
THE IMPACT OF AUTOMATION WILL VARY AMONG ECONOMIES DUE

TO DIFFERENT DEMOGRAPHICS, WAGE LEVELS, PRODUCTIVITY, AND
GROWTH ASPIRATIONS
While automation at a global level can be a significant economic growth engine by
helping reduce the growth gap created by a shrinking workforce, its effects will play out
differently in different countries, depending on the demographic situation in each country,
current productivity levels, wage levels, GDP growth aspirations, and adoption scenarios
for automation.
As we have seen, countries with high shares of aging populations, such as Germany and
South Korea, will have a more urgent need for the productivity boost that automation can
provide. But countries with a range of demographic profiles and growth aspirations also
face challenges and opportunities. We identify three groups of countries, each of which
may be able to use automation to further national economic growth goals. The groups are
advanced economies with aging or shrinking workforces, emerging economies with aging
workforces, and younger emerging economies whose workforces are still growing and
which have high aspirations for future economic growth.
Advanced economies with shrinking workforces could benefit rapidly from

automation to compensate for demographic pressures on growth
Countries including Australia, Canada, France, Germany, Italy, Japan, South Korea, the
United Kingdom, and the United States, with an aggregate annual GDP growth aspiration
until 2030 of 1.8percent, stand to benefit from automation in the medium term. These
advanced economies face a deficit of full-time equivalents even to maintain current GDP
per capita because of their aging populations. Automation could provide enough of a
productivity boost for them to achieve projected GDP for the next few decades.
These countries have considerable incentives to accelerate the pace of automation by
investing in research and development, encouraging the development of hardware and
software solutions to integrate automation technology advances, and other measures that
could make automation more economically feasible. This could more than compensate for
the demographic changes that would otherwise likely slow down economic growth.
Exhibit29 uses the United States as an example of this group. It shows that automation
could make a significant contribution to the productivity increase that the United States
needs to achieve GDP per capita growth. The United States faces a shortfall of around
15million workers already by 2020 just to maintain current GDP per capita. Automation, if
adopted early enough, could enable the United States not only to maintain GDP per capita,
but also to attain the GDP per capita growth rate projected for the economy for the next
several decades, despite the aging population.
85
95
Exhibit 29
In the United States, automation can help achieve projected GDP per capita growth
Projected FTEs
Automation output, earliest
Number of FTEs to achieve projected GDP per capita
Automation output, latest
Number of FTEs to maintain current GDP per capita
Full-time equivalents (FTEs)

Million
400
350
300
250
200
150
100
50
0
1965 70
75
80
85
90
95 2000 05
10
15
20
25
30
35
40
45
50
55
60 2065
NOTE: The projected GDP per capita scenario for the United States uses projections from McKinsey's Global Growth model, with GDP compound annual
growth rate (CAGR) for 201565 of 1.9%, resulting in a productivity CAGR of 1.5%. The maintain current GDP per capita scenario assumes GDP will grow
at the same rate as population (0.5% CAGR for 201565), resulting in a productivity CAGR of 0.1%. See technical appendix for details.
SOURCE: The Conference Board Total Economy database; International Labour Organisation; United Nations Population Division; McKinsey Global Institute
analysis
Emerging economies with aging workforces will get a productivity boost

from automation but will also need to find additional sources to maintain their
growth trajectory
This category includes Argentina, Brazil, China, and Russia, which all face economic growth
gaps as a result of projected declines in the working population. Their first distinguishing
factor is that their current productivity is not sufficient to support GDP per capita over
the long run. For these economies, automation could provide the productivity injection
necessary just to maintain GDP per capita. However, to achieve a faster growth trajectory
that is more commensurate with their developmental aspirations (GDP growth of 4.1percent
and GDP per capita growth of 3.8percent), these countries would need to supplement
automation with additional sources of productivity, such as process transformation and
other technologies.
This grouping is not monolithic, and there are some divergences based on demographic
and economic growth differences. Argentina, Brazil, and Russia, for example, are projected
to have lower GDP per capita growth than China. Argentina and Brazil have younger
populations than China and Russia; the median age is about seven years lower. They also
have significantly faster-growing populations.
While these economies could receive a strong productivity boost from automation, their
wage levels are lower than in the advanced economies we discussed above, and their
overall adoption of automation may be slower as a result, since the business case for
adoption may be less compelling. As we have seen, in these countries it could take longer
for the cost of automation solutions, especially if they involve hardware, to make automation
feasible when compared to the costs of human labor.
96
Policy measures including encouraging increased competition and the development of

a high standard of technological literacy in the population at large could help speed the
process of automation adoption. China, for example, already has the highest rate in the
world for technology-enabled payment platforms; a recent report by market research firm
Nielsen found that 86percent of Chinese paid for online purchases with digital payment
systems, double the global average.86
A second distinguishing factor for these economies that will affect the impact of automation
is that their overall productivity levels tend to be relatively low compared with those of
advanced economies. To capture the full multiplier effects from automation, and achieve
their aspirations for a continued fast-growth trajectory, these countries will need to
supplement automation with other levers to enhance productivity.
Exhibit30 shows the example of China from this group of fast-growing emerging
economies. Chinas population is aging rapidly, which means that in the longer term its
working-age population will peak as early as 2024 and could shrink by one-fifth.87 Within
the next decade, its workforce will be short of the equivalent of about 600million workers
to attain the projected growth if it does not drive productivity improvements. Early adoption
of automation could lessen this gap by about 100million FTEs, but the country still faces a
likely shortfall.
Exhibit 30
Automation can contribute to productivity growth in China, but its high projected GDP per capita growth requires
additional productivity levers
Projected FTEs

Million
6,000
5,000
4,000
3,000
2,000
1,000
0
1965 70
75
80
85
90
95 2000 05
10
15
20
25
30
35
40
45
50
55
60 2065
NOTE: The projected GDP per capita scenario for China uses projections from McKinsey's Global Growth model, with GDP compound annual growth rate
(CAGR) for 201565 of 4.0%, resulting in a productivity CAGR of 4.5%. The maintain current GDP per capita scenario assumes GDP will grow at the same
rate as population (-0.2% CAGR for 201565), resulting in a productivity CAGR of 0.3%. See technical appendix for details.
analysis
China maintains robust e-commerce growth, Nielsen, March 2016

86
87
97
In Exhibit31, we highlight Brazil as another example in this group. The model shows that
early adoption of automation could allow Brazil to meet its medium term GDP per capita
growth expectations. But finding additional levers to accelerate Brazils productivity growth
would be beneficial in both the medium term, as the majority of the adoption scenarios show
automation not providing sufficient economic growth to meet projected GDP per capita
growth, and in the long term.
Exhibit 31
Automation of existing activities could help Brazil meet its medium-term GDP per capita growth aspirations,
but additional productivity acceleration could be required
Projected FTEs

Million
500
450
400
350
300
250
200
150
100
50
0
1965 70
75
80
85
90
95 2000 05
10
15
20
25
30
35
40
45
50
55
60 2065
NOTE: The projected GDP per capita scenario for Brazil uses projections from McKinsey's Global Growth model, with GDP compound annual growth rate
rate as population (0.2% CAGR for 201565), resulting in a productivity CAGR of 0.1%. See technical appendix for details.
analysis
Emerging economies with younger populations will get a boost from

automation but will need other productivity gains to ensure sufficient longerterm growth
The third grouping we identify are emerging economies with younger populations. These
include India, Indonesia, Mexico, Nigeria, Saudi Arabia, South Africa, and Turkey. Saudi
Arabia is something of an anomaly in this group because of its high wages adjusted for
purchasing power parity (see Box6, With a combination of strong population growth and
high wages, Saudi Arabia is atypical). These countries all have aspirations for high longterm growth, in order to lift living standards for the rising population. At the same time, these
countries have strong population growth rates. The ratio of working to total population will
peak in the 2050s (except in Turkey and Saudi Arabia, where it will peak early), ensuring
they have the necessary workforce to maintain GDP per capita. Even so, in these countries,
automation on its own will not suffice to meet the growth aspirations, and other productivity
levers will be needed. Exhibit33 shows how the productivity impact of automation could
play out in Nigeria, which has the highest population growth rate of the group. Still, to meet
the projected GDP per capita growth, it faces a shortfall of ten million workers in 15years.
Early adoption of automation could close this gap.
98
Box6. With a combination of strong population growth and high

wages, Saudi Arabia is atypical
Saudi Arabia and some of its neighbors in the Middle East are experiencing
relatively slow growth, but they have both high wages, measured at
purchasing power parity, and a growing workforce. Saudi Arabia, for example,
has a demographic bulge: more than half the Kingdoms population is younger
than 25, and by 2030 the number of Saudis aged 15years and over will likely
increase by about six million. Based on historical trends in participation, this
upcoming demographic bulge could almost double the size of the Saudi labor
force.1 Its productivity growth of 0.8percent between 2003 and 2013 was well
below the average for its G20 peers.2
Automation could provide a considerable productivity boost to these
countries, enough to meet GDP per capita growth aspirations. Moreover,
automation will be economically feasible in these economies rapidly, because
of their relatively high wage levels. The challenge for governments in these
countries will be to create additional human jobs to employ the large cohort of
young people who will reach working age in the near future. However, the ratio
of working-age population to total population will peak in the 2030s, reversing
the previous demographic dividend and making continued productivity gains
more important (Exhibit32).
Exhibit 32
Saudi Arabia's relatively high wage rates could drive early adoption, and its relatively young demographics
put a premium on creating jobs
Projected FTEs

Million
35
30
25
20
15
10
5
0
1965 70
75
80
85
90
95 2000 05
10
15
20
25
30
35
40
45
50
55
60 2065
NOTE: The projected GDP per capita scenario for Saudi Arabia uses projections from McKinsey's Global Growth model, with GDP compound annual growth
rate (CAGR) for 201565 of 2.2%, resulting in a productivity CAGR of 1.6%. The maintain current GDP per capita scenario assumes GDP will grow at the
same rate as population (0.6% CAGR for 201565), resulting in a productivity CAGR of 0.1%. See technical appendix for details.
analysis
Saudi Arabia beyond oil: The investment and productivity transformation, McKinsey Global
Institute, December 2015.
2
Ibid.
1
99
Exhibit 33
In Nigeria, automation can contribute substantially to economic growth, but adding other means of productivity
growth will likely be required to meet expectations
Projected FTEs

Million
700
600
500
400
300
200
100
0
1965 70
75
80
85
90
95 2000 05
10
15
20
25
30
35
40
45
50
55
60 2065
NOTE: The projected GDP per capita scenario for Nigeria uses projections from McKinsey's Global Growth model, with GDP compound annual growth rate
rate as population (2.4% CAGR for 201565), resulting in a productivity CAGR of -0.5%. See technical appendix for details.
analysis
AUTOMATIONS IMPACT ON EMPLOYMENT COULD BE OF THE SAME ORDER

OF MAGNITUDE AS PREVIOUS MAJOR STRUCTURAL ECONOMIC SHIFTS
A recurring question about automation is its effect on employment. Many forecasters paint
a sometimes dire picture of what the adoption of automation could do for jobs, particularly
blue-collar jobs. The World Economic Forum, for example, has predicted that more than five
millions jobs could be lost to robots in 15 major developed and emerging economies over
the next five years.88
The advent of large-scale automation in the workplace will undoubtedly alter the nature of
the workplace, and the nature of work itself, as machines increasingly take over activities
that were hitherto the domain of human workers. But to some extent, this is an old story.
Telephone operators in the 1950s needed to physically connect switchboard plugs, but
these jobs no longer exist today, as no physical capabilities are required to connect calls.
Telex and telegraph operators are a dying profession. The typing pool that was a mainstay
of office life in the 1950s and 1960s has not survived. In their place have come myriad new
jobs born of the technological developments, from call center employees to IT help desk
personnel. Personal assistants no longer take dictation, but monitor email.
The future of jobs: Employment, skills, and workforce strategy for the fourth Industrial Revolution, World
Economic Forum, January 2016. Other forecasters have made specific predictions about large-scale
job losses related to automation or painted scenarios of a world without work. See Jeremy Bowles, The
computerization of European jobs, Bruegel, July 2014; Martin Ford, Rise of the robots: Technology and the
threat of a jobless future, Basic Books, 2015; and Carl Benedikt Frey and Michael A. Osborne, The future of
employment: How susceptible are jobs to computerisation? Oxford Martin School, September 17, 2013.
88
100
Such anecdotal evidence is underscored by macroeconomic data. Positive gains have been
reported in both productivity and employment in the United States in more than two-thirds
of the years since 1929 despite the rapid onward march of technological development.89
One-third of new jobs created in the United States in the past 25years did not exist, or
barely existed, 25years ago.90 Moreover, every three months about 6percent of US jobs are
destroyed by shrinking or closing businesses, while a slightly larger percentage of jobs are
added.91 This is not just a US phenomenon; findings from other countries indicate that these
are global trends. For example, a detailed analysis of the French economy by McKinseys
French office, published in 2011, showed that while the internet had destroyed 500,000
jobs in France in the previous 15years, it had created 1.2million others, a net addition of
700,000, or 2.4 jobs created for every job destroyed.92
Will this pattern continue with automation, or could things be different this time? (See Box7,
Is this time different?). Certainly the scale and potential scope of the work activities that
have the potential to be automated are very substantial indeed. Our model contemplates the
possibility that hundreds of millions of workers will have to shift the activities they are paid
to do.
Yet even this sort of large-scale structural economic shift over such a long period of time
is not unprecedented. In the United States, for example, the share of farm employment fell
from 40percent in 1900 to 2percent in 2000, while the share of manufacturing employment
fell from 25percent in 1950 to less than 10percent in 2010 (Exhibit34).93 In both cases, the
jobs that disappeared were offset by new ones that were created, although what those new
jobs would be could not be ascertained at the time.
Whatever the net impact on employment, the nature of work will change with automation.
There will be tighter integration between humans and machines than there is today, and this
will increase overall efficiency, since machines will be more accurate with the activities that
they take on. This will free up humans to perform tasks that use higher-level capabilities,
especially those that require social and emotional ones.
Growth and renewal in the United States: Retooling Americas economic engine, McKinsey Global Institute,
February 2011.
90
91
Artificial intelligence, automation, and the economy, Executive Office of the President, December 2016.
92
Impact dinternet sur lconomie franaise: Comment internet transforme notre pays (The internets impact on
the French economy: How the internet is transforming our country), McKinsey & Company, March 2011.
93
Stanley Lebergott, Labor force and employment 18001960, in Output, employment, and productivity in
the United States after 1800, Dorothy S. Brady, ed., NBER, 1966; World Bank data; Mack Ott, The growing
share of services in the US economydegeneration or evolution? Federal Reserve Bank of St. Louis Review,
June/July 1987.
89
101
Box7: Is this time different?

Automation of human work activities has been occurring for at least the past
two centuries, and while these technologies have automated a wide range
of activities people had formerly been paid to do, new activities, occupations
and jobs have been created. This has offset what would have become a
situation of mass unemployment, had new demands for human labor not
been produced.
But is this time different? Are we reaching a point where the pace and/or types
of automating work activities will outstrip the global economys ability to create
new activities and jobs for people to be paid to do?
In some ways, this is an evergreen issue, in that these questions and
concerns have accompanied the adoption of automation through history.
Beyond the time of the Luddites in 19th-century Britain, John Maynard Keynes
wrote about the new disease of technological unemployment during the
Great Depression.1 Over the years, many have speculated about decreasing
demand for human labor as machines automate more work, including Keynes,
in the same article from 1930 quoted above, and the 1966 US Commission
on Technology, Automation and Economic Progress. But while the workweek
has declined from working 10-18 hours per day, six days per week, during the
Industrial Revolution to about eight hours per day, five days per week in the
mid-20th century, it has not declined substantially since that point in many
developed countries.
Some argue that there are factors suggesting that this time is different. Eric
Brynjolfsson and Andrew McAfee describe an inflection point between
the first machine age, based on the automation of physical tasks through
mechanization, and a second machine age, based on the automation
of cognitive tasks through digital technologies.2 Digital technologies
basic capabilities, including computing power, storage capacity, and
communications throughput, appear to be developing exponentially. For
example, Moores Law suggests that the computing power that can be
purchased for $1 doubles roughly every two years. Chroniclers of exponential
technologies such as Ray Kurzweil extrapolate out to a time when a
computer will have the computing power of a human brain, and beyond,
potentially pointing to a future in which, combined with the appropriate
software, an artificial general intelligence could be created that rivals that of
human beings. Others point to a dimension of human work described by the
economists Daron Acemoglu and David Autor: between routine and nonroutine work.3 While much of the work that had historically been automated
was routine (for example, what we describe in this report as predictable
physical activities and the collection and processing of data), many of the
examples of newer automation technologies that we find remarkable automate
what we would have described as non-routine work. That includes driving cars
in busy streets, or diagnosing disease.
John Maynard Keynes, Economic possibilities for our grandchildren, in Essays in

Persuasion, Macmillan, 1933. The essay is available online at www.econ.yale.edu/smith/
econ116a/keynes1.pdf.
2
Erik Brynjolfsson and Andrew McAfee, The second machine age: Work, progress, and
prosperity in a time of brilliant technologies, W. W. Norton & Company, 2014.
3
Daron Acemoglu and David H. Autor, Skills, tasks, and technologies: Implications for
employment and earnings, in Handbook of Labor Economics, volume 4B, David Card and
Orley Ashenfelter, eds., Elsevier, 2011.
1
102
Box7: Is this time different? (continued)

However, the labor market has until now always adapted to the replacement of
jobs with capital, with price effects tending to balance the forces of automation
and creating new complex tasks for people to be paid to do. The December
2016 US White House report on artificial intelligence and automation states:
Recent research suggests that the effects of AI on the labor market in the
near term will continue the trend that computerization and communication
innovations have driven in recent decades The economy has repeatedly
proven itself capable of handling this scale of change, although it would
depend on how rapidly the changes happen and how concentrated the losses
are in specific occupations that are hard to shift from.4 To support some of
the assumptions in our modeling analyses, we argue that many of the factors
affecting the pace and extent of automation adoption, such as the engineering
of solutions to specific problems, and particularly the non-technical
organizational change management, regulatory and acceptance dynamics
around technology adoption, have not changed.
However, changing other assumptions in our model could lead to significantly
different outcomes. For instance, our assumptions on the time required to
develop capabilities and integrate and customize them into solutions that solve
specific problems assumes that these activities will primarily be performed
by people. But in a world where machines can teach themselves, these time
frames could be considerably different from our assumptions.
We also only analyze the automation of current work activities at the
performance levels at which humans currently perform them. If machines
can perform these tasks at significantly higher levels, and/or other new
productive activities, then productivity growth could accelerate even faster
than we modeled. Similarly, we modeled workers displaced by automation
returning to the workforce at the productivity of workers in 2014; if they return
at higher levels of productivity (for example, at the average level of productivity
of workers at the time they are displaced), then overall productivity growth will
also accelerate. Could this lead to a surplus of labor? This would depend on
the economys ability to create new and more things to pay people to do.
Is this time different? We cant definitely saybut the question is a familiar one.
Artificial intelligence, automation, and the economy, Executive Office of the President,
December 2016.
103
Exhibit 34
Employment in agriculture has fallen from 40 percent in 1900 to less than 2 percent today
Distribution of labor share by sector in the United States, 18402010
%
90
Rest of the economy
80
70
60
50
40
30
20
Manufacturing
10
0
1840 50
Agriculture
60
70
80
90 1900 10
20
30
40
50
60
70
80
90 2000 2010
SOURCE: Stanley Lebergott, Labor force and employment 18001960, in Output, employment, and productivity in the United States after 1800, Dorothy S.
Brady, ed., NBER, 1966; World Data Bank, World Bank Group; FRED: Economic Research, Federal Reserve Bank of St. Louis; Mack Ott, The
growing share of services in the US economydegeneration or evolution? Federal Reserve Bank of St. Louis Review, June/July 1987; McKinsey
Global Institute analysis
AUTOMATION COULD CHALLENGE SOME CONVENTIONAL BELIEFS ABOUT

THE GLOBAL ECONOMY AND PATHS TO PROSPERITY
The advances infrom
automation
DUPLICATE
ESwe have outlined and their potential impact on national
economies could upend some prevailing models of development and challenge ideas about
globalization. Among other possible repercussions, it could skewer conventional wisdom
about the economic advantages of having a high birthrate. Low-cost labor may lose its
edge as an essential development tool for emerging economies, as costs of automation
fall. And these economies, in turn, will have new opportunities to leapfrog into higher-value
manufacturing and servicesincluding ITthat will enable them to compete head-on
with advanced economies. Automation could also accelerate the diminishing of trade of
physical goods that began a decade ago, as digital cross-border flows consolidate their
pre-eminence.
Automation could reverse a demographic dividend for economies
The country scenarios we have outlined above could overturn ideas about the relationship
between economic growth and population growth. The declining fertility rates and aging
trends that have taken hold in a broad range of countries including China, Germany, Italy,
and Russia have long been viewed as harbingers of weaker economic growth in the future
and a major policy challenge. The aging and shrinking of the workforce is unprecedented in
modern history, and one of its consequences is that the number of retirees will likely grow
more than twice as fast as the labor pool, leaving fewer workers to support the elderly.94
Conversely, countries with a high birthrate and a growing working-age population were
viewed as having an opportunity to achieve more rapid GDP growth, as they could build
future growth on the twin pillars of productivity and employment growth.
The world at work: Jobs, pay, and skills for 3.5billion people, McKinsey Global Institute, June 2012.
94
104
Automation could help change that scenario. Countries experiencing population declines
or stagnation will be able to make use of it to help maintain living standards even as the
labor force wanes. Meanwhile, countries with high birthrates and a significant growth in
the working-age population may have to worry about where new jobs will come from in a
new machine age. This is especially the case for high-income and low-growth countries,
including some in the Middle East.
The stakes are high because of the sheer numbers of people who have lifted their living
standards as a result of countries following the classic developmental pathand the millions
more who still hope to do so. In India, the impact of automation on employment could affect
working hours of the equivalent of 220million people; in China, that rises to 380million fulltime equivalents.
Chinese companies still have low levels of automation overall, although they are moving to
ramp up. There are only 36 robots per 10,000 Chinese manufacturing workers, about half
the average of all advanced economies and about one-fifth the US level.95 Auto factories
are less than 30percent as automated as US plants, and food processing is only about
12percent as automated as US food processing.96 This gap reflects the cost of Chinese
manufacturing labor, which has risen but remains low by the standards of advanced
economies. The average manufacturing worker makes about 10percent of the average US
manufacturing wage, for example. Our research finds that most Chinese manufacturers
are not yet able to realize the maximal value from robots due to a production process that is
less than optimal. Chinese companies may adopt a hybrid model that mixes the speed and
precision of automation with the flexibility of human labor.
Low-cost labor: No longer a development panacea?

Starting already in the 1880s in Japan, country after country around the world including
South Korea, Taiwan, and most notably China, has followed a familiar pattern of
development. A combination of low-wage agriculture and manufacturingat times often
backed by protectionist policies to encourage import substitution and boost exports
creates jobs and swells household income. As workers become more productive and
households more prosperous, manufacturing moves up the value chain, producing higher
quality products. Country dwellers flock to cities to join this industrialization wave, creating
urban pockets of consumers with disposable income that helps generate greater prosperity.
This trend has been driven by a huge influx of 1.2billion people joining the global labor
market between 1980 and 2010, and it has brought millions out of poverty. In 1990, about
23percent of the worlds population belonged to the consuming class, by which we mean
that they earned more than $10 per day. In 2010, that share had risen to 36percent of the
global population, and we project it will exceed 50percent by 2025.97
That labor-intensive economic development model is still largely intact and is being followed
by countries from India to South Africa. For countries still stuck in poverty, including in subSaharan Africa, it remains the obvious path to prosperity. But in the new world that is taking
shape, low-cost labor may lose some of its edge as an essential developmental tool for
countries, as automation drives down the cost of manufacturing globally. Indeed, research
by the Harvard economist Dani Rodrik suggests that a premature deindustrialization is
already taking place in some emerging economies, although he ascribes that more to trade
and globalization than to technological progress.98
The China effect on global innovation, McKinsey Global institute, October 2015.
Ibid.
97
98
Dani Rodrik, Premature deindustrialization, NBER working paper number 20935, February 2015.
95
96
105
As we have seen, the type of activities most susceptible to automation includes physical
activity or operating machinery in a predictable settingin other words, highly routine
work. A number of occupations in manufacturing sectors that use low-cost labor fit into this
category, such as sewing machine operators, who have a 98percent automation potential.
Agriculture, too, has a high automation potential because much of its activity is both physical
and predictable, and thus replaceable by machines. The speed of adoption of automated
technology depends in part on the size of the firms. For example, in India, where much
farming is on a small-scale family subsistence model, changing to larger-sized farms would
sharply raise the automation potential.
For advanced economies that have lost manufacturing jobs over the past decades because
of competition from lower-cost labor and the buildup of supply chains elsewhere, the drive
to automation may end or even reverse the outflow. The wage gap between advanced
and emerging economies is wider in manufacturing than for other sectors. At the same
time, companies in advanced economies often have a greater ability to fund the capital
expenditure that is needed to build highly automated manufacturing plants.
That does not mean jobs will flow back in large numbers, since only a very small part of
manufacturing value added is driven by labor cost, and any re-onshoring that does take
place will likely happen in highly automated plants. Moreover, for companies in advanced
economies, the business case for automation is not simple: hardware and software can
be costly, and integrating them successfully is laborious. But machines also have some
clear advantages: they can run continuously and do not require large human resources
departments. Operations are easier to manage if they are next door, which makes oversight
and shipping easier than if operations were halfway around the world. Politically and from a
public relations standpoint, too, it can be advantageous to be seen as a local stalwart.
New opportunities for higher-value manufacturing and services

through automation
Growing use of automation across sectors and within sectors raises the prospect that
some countries could leapfrog in the future to become active in industries where they are
now weak or have little presence and either no infrastructure base or one that is aging and
technically obsolete. For example, Saudi Arabia and Iran have aspirations to build up an
automotive industry to serve not just the domestic market but also the wider Middle East
region. They both have domestic supplies of raw materials, such as iron ore and bauxite,
and plentiful energy; deploying automation technologies could help them leapfrog into stateof-the-art manufacturing. In Russia, where the number of employees will likely decrease by
30percent over the next half century as a result of a declining birthrate, automation could
compensate for the smaller workforce and revitalize growth in manufacturing sectors.
New technologies and integrated solutions might reduce capital expenditure, benefiting
emerging economies with limited resources. They also potentially provide the ability for
emerging economies to create economies of scale.
Emerging economies could gain an edge in some service sectors, including health care
and social assistance, IT, and professional, scientific, and technical services. That is
because the gap in wages between countries for these services is less pronounced than it
is for manufacturing, and the cost of the software needed to compete on a global scale is
much lower.
106
Automation could accelerate the diminishing of physical trade

What we now can see as the heyday of globalization, the 20-year period that began in the
mid- to late 1980s, the global flows of goods, services, and finance grew rapidly, outpacing
GDP growth. Since the global financial recession in 2008 and slow recovery, however, that
rapid expansion has stopped in its tracks. Growth in global goods trade has flattened, trade
in services has posted only modest growth, and financial flows have fallen sharply. At the
same time, data flows have soared, with cross-border bandwidth growing 45 times larger
since 2005. Data flows now account for a larger impact on of global GDP than does global
trade in goods.99
Automation could to some extent push this trend further, as companies rely less on
traditional shipping methods and more on digital transactions and exchanges. For example,
3D printing, if widely adopted by global manufacturers, could reduce global trade volumes
as more products are printed where they are consumed. There are already examples
of this at work, such as GE Aviation, which is beginning to use 3D printing to produce fuel
nozzles for its new Leap engine.100 A fuel nozzle made the traditional way consists of 20
components, with a supply chain that spans countries. 3D printing allows the company
to produce best-quality nozzles in one piece, at one location, eliminating the need to ship
intermediate parts across borders.
At the same time, the rise of digital platforms such as Alibaba, Amazon, and eBay changes
the economics of doing business across borders, bringing down the cost of international
interactions and transactions. These platforms create markets and user communities
with global scale, providing businesses in advanced and emerging economies alike with
a huge base of potential customers and effective ways to reach them. Companies based
in developing countries can overcome constraints in their local markets and connect with
global customers, suppliers, financing, and talent far more easily than they ever could.101
Automation can quickly become a new engine for the global economy at a time when the
working-age population in numerous countries is stagnant or falling and productivity growth
is struggling to compensate. Whatever their economic structure, wage levels, growth
aspirations, or demographic trends, countries around the world could benefit from adopting
automation to maintain living standards and help meet long-term growth aspirations. The
rapid development and growing adoption of automation technologies will create myriad
new opportunities even as they likely disrupt the world of work and challenge long-held
conventions about the global economy and paths to prosperity. In order to make the most
of the potential offered by automation and, at the same time, manage its consequences on
companies, national economies, and workers around the world, policy makers, business
leaders, and men and women everywhere will need to think through the implications that
these new technologies will bring and prepare for significant changes. In our final chapter,
we discuss how stakeholders around the world can position themselves to benefit fully from
automations potential while avoiding its pitfalls.
Digital globalization: The new era of global flows, McKinsey Global Institute, February 2016.
Ibid.
101
Playing to win: The new global competition for corporate profits, McKinsey Global Institute, September 2015.
99
100
107
Joos Mind/Photographers Choice/Getty Images

108
6. PREPARING FOR DISRUPTION

Automations challenge for policy makers, business leaders, and workers everywhere is a
formidable one: how to capture the positive effect on the global economy, at the same time
as navigating what is likely to be a complicated period ahead, one with potentially epochal
social, economic, and employment repercussions.
At a time of sluggish GDP growth and weak productivity gainsand when demographic
trends are starting to work against growth in a broad range of countriesautomation
could serve as an unforeseen boon to the world economy. Yet anxieties about lost jobs
and reduced incomes are already creating a backlash against globalizing and modernizing
trends, especially in advanced economies, influencing election outcomes in several
countries. Ever since the Industrial Revolution, evolving technologies have aroused fears as
well as excitement. The risk that automation could become a scapegoat is real.
As already noted, the public debate over automation takes place against a backdrop of a
growing gap in incomes and employment prospects of high-skill and low- and middle-skill
workers.102 The share of national income that is paid to workers, the so-called wage share,
has been declining in many advanced economies even as productivity has risen, suggesting
a disconnect between productivity and incomes, which automation could potentially
exacerbate further. The wage-share decline is due in part to the growth of corporate profits
as a share of national income, as a result of rising capital returns to technology investments,
lower returns to labor from increased trade, rising rent incomes from homeownership, and
increased depreciation on capital.103
Moreover, there is already a significant mismatch of skills in the global workforce, with high
levels of youth unemployment and, at the same time, a shortage of job seekers with critical
skills. Overcoming this mismatch is a complex undertaking that requires close cooperation
among education providers, governments, and businesses.104
Uncertainty about the timing of automation adoption and its potentially variable impact
from sector to sector, from country to country, and from workplace to workplace, make
the challenge of preparing for it even more complex. Yet preparation is both possible and
necessary, within the business world, at a policy level, and for individuals. Automation
technologies are advancing rapidly, and those who harness them effectively and take the
lead in their sectors will gain a competitive advantage.105 It is never too early to think through
strategic options and appropriate responses.
See the discussion of technical change and skills in Chapter 4.

While overall spending on capital goods has been weak, there has been considerable investment in
information technology, whose prices have declined. See Loukas Karabarbounis and Brent Neiman, The
global decline of the labor share, NBER working paper number 19136, June 2013; Loukas Karabarbounis
and Brent Neiman, Declining labor shares and the global rise of corporate saving, NBER working paper
number 18154, June 2012; and How CBO projects income, Congressional Budget Office (CBO), July 2013.
See also, Poorer than their parents? Flat or falling incomes in advanced economies, McKinsey Global Institute,
July 2016.
104
For a detailed discussion of the education to employment transition, see Education to employment: Designing
a system that works, McKinsey Center for Government, January 2013. See also, A labor market that works:
Connecting talent with opportunity in the digital age, McKinsey Global Institute, June 2015.
105
For details of how companies that are leaders in digitization increase profit margins and raise productivity, see
Digital America: A tale of the haves and the have-mores, McKinsey Global Institute, December 2015.
102
103
AUTOMATION WILL PROVIDE SIGNIFICANT OPPORTUNITIES FOR COMPANIES

TO CAPTURE VALUE, BUT IT WILL HEIGHTEN COMPETITION AND REQUIRE
CHANGED BUSINESS ROLES AND WORKFLOWS
Capturing the full opportunities offered by automation including both the labor substitution
and other performance benefits is likely to give companies a competitive advantage, but
doing so will likely require them to conduct a thorough review of corporate activities and
potential overhaul of business processes and workflows.
Automation will enable new forms of competitive advantage, but it will also
require companies to raise their game just to keep up
There are many opportunities for companies to take advantage of the potential of
automation to seek competitive advantage. Automation of various activities can improve the
performance of almost any business process, as we noted in our case studies in Chapter3.
Certainly, automation can be used to transform the costs of a process by reducing
labor costs, for example when end-to-end digitization is used to create straight-through
processing of a transactional process. As we have also documented, automation can not
only enable a reduction in labor costs, it can also bring a range of other benefits related to
performance improvements, such as greater throughput, improved reliability, raised quality,
better safety, and other gains. Straight-through processing of financial transactions, for
instance, is usually faster than the manual process it replaces, and reduces the number of
errors introduced into the process.
Thus, any of the benefits that automation can unlock could become a basis of competition.
An industrial company that is able to ensure significantly lower downtime than its
competitors by automated monitoring and predictive maintenance of its equipment can
compete on this basis of better reliability. A consumer company that can provide faster
delivery and 24x7 customer service, through automation in its supply chain and contact
centers, can compete on the basis of being more responsive to its customers. Automation
can also enable companies to create new products, services and/or business models,
for example, a professional services company that is able to provide customized advisory
services to small and medium-sized businesses, or even consumers, through an automated
conversational interface.
Some of the strategic capabilities that automation can unlock are more subtle and wideranging than improving the performance of a particular process or offering. Some forms
of automation, for example those that are based on machine learning techniques such as
deep learning, improve their performance over time when they have access to more data.
Companies that are able to create platforms with increasing returns to scale can create
network effects that result in winner-take most dynamic. Machine learning on user data
allows the platform to become more compelling to users, which in turn generates more
data.106 We have seen these types of dynamics in online search platforms, social media
platforms, media delivery platforms, and increasingly platforms that support activities in
the physical world, such as platforms for Internet of Things data, healthcare, and travel
and transportation.
Automation could also unlock the otherwise unlikely combination of scale and agility,
with the ability to instantly propagate changes across an entire organization. When an
organizations activities are controlled by automatic systems, modifying the behavior of the
enterprise can be accomplished by software download, rather than an extensive change
management program. When automation is used to augment human management,
traditional organizational orthodoxies, such as about spans of control, can be challenged.
For example, Uber takes advantage of automation for coordination, and has only about one
106
110
Michael Chui and James Manyika, Competing at the digital edge: Hyperscale businesses, McKinsey
Quarterly, March 2015.
6. Preparing for disruption
human manager per 1,000 drivers compared with a typical limousine company that has
about one manager per 20 to 30 drivers. However, greater scale and speed also means that
the effects of negative changes are amplified. And these are not the only challenges that
automation could bring to businesses. From a technology standpoint, as software and data
underlie nearly all new automation technologies, cybersecurity issues become more critical
than ever.
Strategically, automation can also heighten competition, enabling startups or firms
from other sectors to encroach on new turf and exacerbating a growing divide between
technological leaders and laggards in every sector.107 At the same time automation can
support increasing scale, it also can provide the leverage that enables smaller companies
to compete with much larger firms. When the basis of competition using automation is
algorithms and data, the raw materials of competition are more readily accessible, through
the cloud and open data. For example, we have seen the advent of new competitors in the
media and game sectors deploying automation tools to quickly gain the reach of much more
established players.
The fact that automation displaces human work activities creates a set of real challenges for
companies that deploy it. Worker displacement creates the potential for labor unrest, and
the size and scope of retraining/placement programs, whether put in place by companies or
others, will have to mirror the size and scope of automation programs. The effective use of
automation requires the transformation of processes, changing what people do, even those
that are not made redundant by automation. In general, workflows will change, and new
roles will emerge, such as that of robot trainer or exception handler.
Companies embracing automation could experiment and map areas of likely

impact, even as they focus on workplace changes and new skills for workers
Companies who recognize both the opportunities and threats of automation to
competitiveness will engage and embrace the potential that these technologies represent,
prioritizing a set of active experiments to start climbing the learning curves earlier
rather than later. To help diagnose where automation could most profitably be applied
to improve performance, business leaders may want to conduct a thorough inventory
of their organizations activities and create a heat map of where automation potential
is high. Business processes shown to have activities with high automation potential
could be reimagined under scenarios where they take full advantage of automation
technologies (rather than mechanically attempting to automate individual activities using
current processes). The benefits and feasibility of these automation-enabled process
transformations could then be used to prioritize which processes to transform using
automation technologies.
Business leaders and their organizations will also need to become more knowledgeable
about the evolution of the technologies themselves, understanding the art of the
possible, and the potential for the future, in order to best position their enterprises to take
advantage of automation. This is not just book knowledge that comes from reading
about technologies, or visiting global centers of innovation, but practical knowledge that
comes from devoting some resources to continually and purposefully experimenting with
technologies on real problems, and then scaling those that demonstrate promise.
Perhaps the most vital component to being successful at deploying automation is the hard
work that has to be done to prepare and adapt human capital to work in complementary
ways with technology. As our activity-based analysis has shown, almost every role will
change, and every workflow eventually will be transformed. Workers will have to be
continually retrained as the work activities that they do, to work alongside machines,
This is already happening with digital technologies. See ibid.
107
111
continue to evolve. Others will have to be redeployed, potentially to other positions in the
economy, and businesses have a role to play in aiding these transitions. This will require
not only changes in skills, but also changes in mindsets and culture, in a world where
work activities continue to change, and co-workers include not only other people, but
also machines.
FOR POLICY MAKERS, AN EMBRACE OF AUTOMATION COULD GO HAND IN

HAND WITH MEASURES TO SUPPORT LABOR DEPLOYMENT
Policy makers must recognize the pressing need for productivity acceleration to
compensate for demographic aging shifts in order to enable GDP per capita growth.
Automation technologies can provide a major contribution to accelerating productivity
growth. Thus there are two broad categories of issues for policy-makers to consider. First,
how can we accelerate the development and deployment of automation to generate greater
growth in productivity? Second, how can we support the redeployment to other productive
activities of workers whose activities are automated?
Policy makers can accelerate early development and adoption of
automation technologies
Early adoption of automation could benefit from policy support, both in regard to the
technology development and for its deployment. This support could include investments
in developing the technologies themselves, including funding basic research and support
for commercialization, as well as supporting investments in digitally enabled infrastructure
for automation.
Investment in enabling infrastructure for automation adoption could be an early priority,
especially for emerging economies that may not be as digitally enabled as some advanced
economies. In general, large-scale automation will require substantial investment, and
the tax and other treatment of this investment could enableor hinderthe adoption of
automation technologies. For regulators, automation can pose challenging issues for safety
and liability; for example, in the case of self-driving vehicles, who could be held liable for
accidentsthe automaker, the owner, or the algorithm creator? Thoughtful regulatory dialog
and policy making will be important to ensure that the benefits of automation are achieved
while protecting other societal concerns.
Engaging a broader societal dialog about automation, the need for productivity growth,
and shifts in labor markets is another role that policy-makers can play. Deployment of some
technologies could face concerted opposition from unions or other labor organizations over
concerns about the employment impact. Governments will need to find cogent answers and
coherent policies to engage in these debates
Exposing and stimulating the work that needs doing

Governments are often not particularly able by themselves to anticipate the types of jobs
that could be created, or new industries that will develop (and they are not alone in this
limitation). However, they are well positioned to catalyze dialogues about what work needs
doing, and the grand societal challenges that require more attention and human effort.108
The 1966 US Report of the National Commission on Technology, Automation and Economic
Progress devoted multiple chapters to Unmet Human and Community Needs, including
sections on education, healthcare, urban transportation, air pollution, water resources,
housing, and international development, all of which seem as relevant in this era as they
were 50years ago.109 Perhaps a similar report today would add caring for the elderly to the
list, as we have documented the demographic effects of aging.
Tim OReilly, Dont replace people. Augment them, Medium.com, July 17, 2016.
Technology and the American economy: Report of the National Commission on Technology, Automation, and
Economic Progress, US Department of Health, Education, and Welfare, 1966.
108
109
112
While policy-makers might not be able to predict the new activities and occupations that can
be created, they can help create the conditions under which innovation in the use of human
labor becomes more likely. Governments could also encourage new forms of technologyenabled entrepreneurship. Digital technology itself can enable new forms of entrepreneurial
activity. Workers in small businesses and self-employed occupations can benefit from
higher income earning opportunities. A new category of knowledge-enabled jobs will
become possible as machines embed intelligence and knowledge that low-skill workers
can access with a little training. In India, for example, Google is rolling out the internet Saathi
(friends of the internet) program in which rural women are trained to use the internet and
then become local agents who provide services in their villages through internet-enabled
devices. The services include working as local distributors for telecom products (phones,
SIM cards, and data packs), field data collectors for research agencies, financial-service
agents, and para-technicians who help local people access government schemes and
benefits through an internet-based device. There is a need for these services in rural India
where broad internet reach and digital literacy are still low, but where an increasing number
of services are being provided online. Googles program aims to create more than 50,000
internet Saathis who will provide services to more than 50million households in rural India.110
Addressing wages, skills gaps, and labor market mismatches

One of the challenges of the new era will be to ensure that wages are high enough for the
new types of employment that will be created, to prevent continuing erosion of the wage
share of GDP, which has dropped sharply since the 1970s.111 While some governments
may be tempted to look for ways to slow automation adoption, out of concern for possible
employment effects, such moves could prove counterproductive, holding back productivity
without protecting jobs durably.
Automation could exacerbate a skills gap, even as it touches all occupations. There is
already a growing divide in income advancement and employment opportunities between
high-skill workers and those who are low- and medium-skill. In the past two decades,
there has been a clear pattern of consistent job growth for high-skill workers and little or no
growth for low- and middle-skill ones. For example, in 1981, college-educated workers in
the United States earned a 48percent wage premium over high school graduates. By 2005,
that premium had risen to 97percentin other words, an American college graduate earns
almost twice as much as a high school graduate.112 The growing gap between productivity
and wages is not new, but automation could accelerate the process. In its 2016 report on
automation, the White House noted that the trend toward skill-biased change brought about
by computerization and communications innovations is likely to continue in the decade
ahead as a result of artificial intelligences effects on the labor market.113
To address this gap, policy makers could work with education providers to improve basic
skills through the schools system and put a new emphasis on capabilities that are among
the most difficult to automate, including creativity, understanding human emotions, and
managing and coaching others. For people who are already in the workforce, they could
intervene to help workers develop skills best suited for the automation age. For example,
many economies are already facing a shortage of data scientists and business translators.114
Governments working with the private sector could take steps to ensure that such gaps
Indias technology opportunity: Transforming lives, empowering people, McKinsey Global Institute,
December 2014.
111
Poorer than their parents? Flat or falling incomes in advanced economies, McKinsey Global Institute,
July 2016.
112
David Autor, Skills, education, and the rise of earnings inequality among the other 99percent, Science,
volume 344, issue 6186, May 2014. See also Poorer than their parents? Flat or falling incomes in advanced
economies, McKinsey Global Institute, July 2016.
113
Artificial intelligence, automation, and the economy, Executive Office of the President, December 2016.
114
The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016
110
113
are filled, with new education and training possibilities established rapidly and prioritized.
They could also foster the growth of technology-enabled solutions for the labor market that
improve matching and access to jobs, such as online talent platforms.115 As automation
reshapes the workplace, independent work could become increasingly important, and
policy makers will want to address issues such as benefits and variability that these
platforms can raise.
Furthermore, while important work that needs doing might be identified, there is a possibility
of market failures in the wages that might be paid for that work, a situation for which public
policy interventions might be appropriate.
Rethinking social support

Full or partial automation will result in labor displacement, and it will be important to
support workers as they transition from one set of activities to another. As work evolves at
higher rates of change between sectors, locations, activities, and skill requirements, many
workers may need assistance in adjusting to the new age. This could involve providing
support during transitional periods, for example retraining or income support. While our
modeling suggests a higher likelihood of labor shortages than labor surpluses, there might
be people whose skills and capabilities are mismatched to the work that needs doing, or
where wages are put under pressure by specific increases in labor supply (for example,
within a geography, for workers with particular skills, in specific industries). In these cases,
adapted social safety nets could help provide support. Various ideas have been considered,
including work sharing, negative income taxes, and universal basic income (see Box8,
When some old policy ideas are new again).
FOR WORKERS, AUTOMATION WILL CHANGE MANY WORK PROCESSES AND
REQUIRE A CLOSER COLLABORATION WITH TECHNOLOGY
Regardless of the longer-term implications, in the short to medium term, men and women
in the workplace will need to engage more comprehensively with machines as part of their
everyday activities. Tighter integration with technology will free up time for human workers
including managers to focus more fully on activities to which they bring skills that machines
have yet to master. This could make work more complex, and harder to organize, with
managers spending more time on coaching.
As young people in particular make education and career choices, it will be important for
them to be made aware of the factors driving automation in particular sectors, to help them
identify the skills that could be useful for them to acquire from a labor-market perspective,
and what activities will be complements of activities that are likely to be automated.116
High-skill workers who work closely with technology will likely be in strong demand. Those
involved in developing and deploying automation technologies will have many opportunities.
In addition, workers who are paid to do activities that are complements of automation
will also find themselves in an advantageous position, as Brynjolfsson and McAfee have
described it, racing with the machines rather than racing against the machines.117 These
and other workers may be also able to take advantage of new opportunities for independent
work as the corporate landscape shifts and more project work is outsourced by big
companies. Low-skill workers working with technology will be able to achieve more in terms
of output and productivity but may experience wage pressure given the potentially larger
supply of similarly low-skill workers.
A labor market that works: Connecting talent with opportunity in the digital age, McKinsey Global Institute,
June 2015.
116
Erik Brynjolfsson and Andrew McAfee, The second machine age: Work, progress, and prosperity in a time of
117
Ibid.
115
114
Education systems will need to evolve for a changed workplace, with policy makers working
with education providers to improve basic skills, with a new emphasis on topics such as
creativity, emotional intelligence, and leading and coaching others. For all, developing agility,
resilience, and flexibility will be important at a time when everybodys job is likely to change
to some degree.
Finally, automation will create an opportunity for those in work to make use of the innate
human skills that machines have the hardest time replicating: social and emotional
capabilities, providing expertise, coaching and developing others, and creativity. For
now, the world of work still expects men and women to undertake rote tasks that do not
stretch these innate capabilities as far as they could. As machines take on ever more of the
predictable activities of the workday, these skills will be at a premium. Automation could
make us all more human.
Box8. When some old policy ideas are new again

Many of the potential policy measures that could be
adopted to help the labor force adjust to the impact
of automation are not entirely new. The 1966 US
Commission on Technology, Automation, and Economic
Progress recommended taking actions that included
improving education and training, facilitating better
matching between workers and work (including greater
transparency for workers), creating portable benefits
that follow workers across different jobs, and increasing
work-hour flexibility. These are all ideas that find echoes in
todays discussions around the world.1
Another idea that has returned is providing a universal
basic income, in other words, providing all citizens
with an unconditional sum of money. Automation has
given it a new lease of life among policy makers, some
academics and a number of business leaders in Silicon
Valley, although it remains controversial.2 In a June 2016
referendum, Swiss voters overwhelmingly rejected a
proposal to establish a universal basic income.3
A full basic income program has never been enacted
and properly studied. However, in Finland, an experiment
that started on January 1, 2017, will pay an unconditional
basic income of 560 euros per month for two years to a

random sample of 2,000 individuals drawn from current
working-age beneficiaries of unemployment benefits. The
experiment is aimed at comparing the employment rate of
beneficiaries of the basic income with those who receive
traditional unemployment benefits.4
Others have suggested that if we need human labor
working alongside automation to achieve economic
growth, social assistance programs should incentivize
work, such as negative income taxes. The history of a
negative income tax for low-paid workers spans back
to the 1940s, when it was proposed by British politician
Juliet Rhys-Williams, and it was advocated by Milton
Friedman in the 1960s. In 1975, the United States
introduced a negative income tax, the earned income tax
credit, which provides income subsidies to the working
poor. The program has survived for 40years and today
annual payments range from $500 for an individual
with no children earning less than $14,820, to $6,242
for a family with three or more children and household
income of less than $53,267.5 Other countries have
similar programs.
See, A labor market that works: Connecting talent with opportunity in the digital age, McKinsey Global Institute, October 2016; Independent
work: Choice, necessity and the gig economy, McKinsey Global Institute, October 2016; and The world at work: Jobs, pay and skills for
3.5billion people, McKinsey Global Institute, October June 2012.
2
See for example, Charles Murray, A guaranteed income for every American, Wall Street Journal, June 3, 2016. Among business leaders, Elon
Musk has spoken out in favor of such a program.
3
Swiss voters reject proposal to give basic income to every adult and child, Guardian, June 5, 2016.
4
Preparations for the basic income experiment continue, Kela, December 14, 2016. In the 1970s, Canada launched a five-year experiment in
guaranteed basic income, known as Mincome, in Dauphin, Manitoba. The poverty level decreased, hospitalization rates fell, and high school
completion rates rose. The drawback was that non-primary income earners (often mothers of small children) dropped out of the labor force.
See Evelyn L. Forget, The town with no poverty: A history of the North American guaranteed annual income social experiments, University of
Manitoba, May 2008.
5
US Internal Revenue Service.
1
115

Considerable uncertainties surround the advent of the automation eraand have done
so for years. A half century ago, a US commission on technology, automation, and
economic progress wrestled with some of the same questions about the future of work and
employment that we do in this report.118 The speed with which automation will be adopted
into the workplace will vary, and the effects on employment, on national economies, and
on businesses and workers globally will play out in myriad ways. At its core, however,
automation represents a considerable opportunity for the global economy at a time of weak
productivity and a declining share of the working-age population. For corporate leaders,
too, automation will reshape the business landscape and create considerable future value.
How to capture the opportunities and prepare for the possible consequences will be a key
political, economic, corporate, and social question going forward. This is not something
that can be watched from the sidelines. Automation is already here, and the technological
advances continue. It is never too early to prepare.
Technology and the American economy: Report of the National Commission on Technology, Automation, and
Economic Progress, US Department of Health, Education, and Welfare, 1966.
118
116
In 2016 a computer beat the human world champion at the game of Go.
Steve McAlister/Getty Images
118
TECHNICAL APPENDIX
This appendix outlines key points on the methodology in the following sections:
1. Assessment of technical potential for automation
2. Modeling of automation adoption timelines
3. Economic modeling
4. Key proxies and data sources
1. ASSESSMENT OF TECHNICAL POTENTIAL FOR AUTOMATION

We assess the technical potential for automation of the global economy through an
analysis of the component activities of each occupation. Our analysis covers 46 countries
representing more than 80percent of the global economy. We used databases published
by institutions including the World Bank and US Bureau of Labor Statistics 2014 O*Net
database to break down about 800 occupations into more than 2,000 activities, and we
determined the performance capabilities needed for each activity based on the way humans
currently perform them.
We further break down activity into 18 capabilities and assess the technical potential of
those capabilities. This framework is informed by academic research, internal expertise, and
industry experts. For each capability, we define four possible levels of requirement, ranking
from not required to the equivalent to top-quartile human performance. ExhibitsA1 to A4
give a high-level summary of the capabilities and the four continuum criteria.
Exhibit A1
Criteria for automation: Accept input
Score (based on tech advancements and complexity)
Requirement:
Basic to
execute task
Automation
capability
Requirement:
High human (top
quartile of global
labor pool)
Metric to define
continuum
Natural
language
understanding
Does not require

NLU
Low language
comprehension
required (while still
accurate with
structured
commands)
Moderate language
comprehension
(medium accuracy
of nuanced
conversation)
High language
comprehension
and accuracy,
including nuanced
human interaction
and some quasi
language
Accuracy of
comprehension
Complexity of
language/
context/
integration
Sensory
perception
Does not require

Autonomously
sensory perception infers simple
external perception
(e.g., object
detection, light
status,
temperature) using
sensory data
Autonomously
infers more
complex external
perception using
sensors (e.g., high
resolution detail,
videos) and simple
integration using
inference
High human-like
perception
(including ability to
infer and integrate
holistic external
perception)
Accuracy of
perception/
complexity of
scene
Degree of
integration
across sensors
Social and
emotional
sensing
Does not require

social and
emotional sensing
Comprehensive
social and
emotional sensing
(e.g., voice, facial
and gesture
recognition-based
social and
emotional sensing)
High human like

social and
emotional sensing
Quality of
comprehension
Basic social and

emotional sensing
(e.g., PAD emotion
model in
mechanical
systems)
120
Automation
Technical appendix
Exhibit A2
Criteria for automation: Information processing
Requirement:
Basic to
execute task
Automation
capability
Requirement:
High human (top
quartile of global
labor pool)
Metric to define
continuum
Recognizing
Does not require
known patterns/ pattern/category
category
recognition
(supervised
learning)
Recognition of
basic known
patterns/categories
(e.g., lookup
functions in data
modeling)
Recognition of
more complex
known patterns/
categories
High human-like
recognition
of known patterns
Complexity of
pattern
Generation of
novel patterns/
categories
Simple/basic ability
for
pattern/category
generation
More advanced
capacity for
recognition of new
patterns/categories
and unsupervised
learning
High human-like
recognition of new
patterns/
categories,
including
development of
novel hypotheses
Complexity of
pattern
Logical
Does not require
reasoning/
contextual
problem solving knowledge or
interpretation of
novel inputs
Capable of
problem solving
based on
contextual
information in
limited knowledge
domains with
simple
combinations of
inputs
Capable of
problem solving in
many contextual
domains with
moderately
complex inputs
Capable of
extensive
contextual
reasoning and
handling multiple
complex, possibly
conflicting, inputs
Complexity of
context and
inputs
Optimization
and planning
Does not require

optimization and
planning
Simple
optimization (e.g.,
optimization of
linear constraints)
More complex
optimization (e.g.,
product mix to
maximize
profitability, with
constraint on
demand and
supply)
High human-like
optimization and
planning based on
judgment (e.g.,
staffing a working
team based on
team/individual
goals)
Degree of
optimization
(single vs. multi
variate)
Creativity
Does not require

creativity
Some similarity to
existing
ideas/concepts
Low similarity to
existing
ideas/concepts
No similarity to
existing
ideas/concepts
Novelty/
originality and
diversity of ideas
Information
retrieval
Does not require

information
retrieval
Search across
limited set of
sources (e.g.,
ordering parts)
Search across
multiple set of
diverse sources
(e.g., advising
students)
Expansive search
across
comprehensive
sources (e.g.,
writing research
reports)
Scale (breadth,
depth, and
degree of
integration) of
sources
Speed of
retrieval
Coordination
with multiple
agents
Does not require

collaboration
Limited group
collaboration; low
level of interaction
Regular group
interaction
requiring real-time
collaboration
Complex group
interaction
requiring high
human like
collaboration
Complexity of
coordination
(i.e., number of
interactions per
decision)
Speed/frequency
of coordination
Social and
emotional
reasoning
Does not require

social/emotional
reasoning
Basic social and

emotional
reasoning
More advanced
social and
emotional
reasoning
High human-like
social and
emotional
reasoning
Complexity of
emotional
inference
Does not require

pattern/category
generation
121
Exhibit A3
Criteria for automation: Deliver output
Requirement:
Basic to
execute task
Requirement:
High human (top
quartile of global
labor pool)
Automation
capability
Output
articulation/
display
Does not require

any articulation/
display
Articulation of
simple content
(e.g., organizing
existing content)
Articulation of
moderately
complex content
High human-like
articulation
Complexity of
message
delivered
Variability in
medium of
message
delivered
Natural
language
generation
Does not require

NLG
System output with

basic written NLG
(e.g., web crawl
results)
System output with

advanced NLP
(more complex
structure)
Nuanced, high
human-like
language output
Complexity of
message
delivered.
Note: includes
use of quasi
linguistics
(idioms, common
names, etc.)
Accuracy of
audience
interpretation
Emotional and
social output
Does not require

social and
emotional output
Simple social and

emotional
discussions (e.g.,
conversations with
no gestures)
Advanced social
and emotional
discussions (e.g.,
conversations with
gestures)
Nuanced high
human-like body
language and
emotional display
Complexity of
emotional
communication
Accuracy of
audience
interpretation
122
Technical appendix
Metric to define
continuum
Exhibit A4
Criteria for automation: Physical movement
Requirement:
Basic to
execute task
Requirement:
High human (top
quartile of global
labor pool)
Automation
capability
Metric to define
continuum
Fine motor
skills/dexterity
Does not require

physical dexterity
Ability to handle
and manipulate
common simple
objects (e.g., large
solid objects) using
sensory data
Can handle and

manipulate wide
range of more
complex and
delicate objects
(e.g., pick up egg)
High human
dexterity and
coordination
Precision,
sensitivity, and
dexterity of
manipulation
Gross motor
skills
Does not require

gross motor skills
Basic 1D/2D motor

skills
More advanced
multi-dimensional
motor skills
High human multidimensional motor

skills
Range and
degree of motion
Speed and
strength of
motion
Navigation
Does not require

localization/
navigation
Use pre-defined
algorithm for
mapping and
navigation
Autonomous
mapping and
navigation in
simple
environment
Autonomous
mapping and
navigation in
complex
environment
Complexity of
environment
(while still
maintaining
accuracy)
Mobility
Does not require

mobility
Mobility/locomotion
in simple
environment
(e.g., limited
obstacles/office
space)
Mobility/locomotion High human

mobility
in more complex
and locomotion
terrain of human
scale environment
(e.g., climbing
stairs)
Speed (gross
motor) of
mobility
Scale of mobility
(2D vs. 3D)
Complexity of
environment/
terrain
Assigning the required level of capabilities to activities

We used a machine-learning algorithm to score the more than 2,000 work activities in
relation to the 18 performance capabilities. To train the algorithm we devised a list of
keywords that we shared with experts. The algorithm scores each activity by matching
keywords from the capability to the activity title. Where we found anomalies, special
requirements, or a need for nuance, we made adjustments, for example, in assessing the
level of capabilities needed to navigate in extreme weather or on uneven surfaces and other
unpredictable settings, or the different physical capabilities required by a kindergarten
teacher compared with a middle school teacher.
123
2. MODELING OF AUTOMATION ADOPTION TIMELINES

Our adoption model assesses the automation development and adoption timeline at activity
level for more than 800 occupations across 19 sectors as well as 46 countries, which
represent more than 80percent of the global economy. We divide the adoption process
into four phases: technical feasibility, solution development, economic feasibility, and enduser adoption.
For a work activity to be automated, every performance capability needed to carry out that
activity must advance to the required technical level. To develop progression scenarios
for the development capabilities, we used survey findings, extrapolation of metrics, and
other predictors of technical advances. We conducted interviews with industry leaders and
academic experts. We also looked at some recent commercial successes showcasing
capabilities, as well as historic trajectory of capabilities. We then adjusted the result using
some identified constraints that could limit the progression of certain capabilities.
To improve accuracy, we looked at other predictors, including from the press, companies,
research groups, and relevant industries. These predictors include research and
technological breakthroughs, trends in publications, and patents as a measure of research
potential. We also looked at predictors of computational resources that are required for
capabilities as compared with Moores law. Another important predictor is open source and
crowdsourcing developments that could potentially accelerate the technology growth.
Based on the assessment, we projected the reasonable expected time frames to reach the
next level of performance for each capability.
To estimate how long it would take to develop a solution that could integrate automation
technologies after the establishment of technical feasibility, we used a three-step process.
First, to understand the solution development life cycle, we collected the development time
and technical capabilities for more than 100 previously developed automation solutions,
including both hardware and software solutions. We recorded the number of years from
initial research to development (product launch) as well as up to three of the most relevant
capabilities for each.
Second, we generated solution development scenarios, taking the 25th and 75th percentile
development time for each of the 18 capabilities across all of its associated solutions.
Third, we estimated the solution development time for a given work activity based on the
maximum solution development time of capabilities that activity required. Several examples
are assessed to ensure accuracy, and we assign 25th and 75th percentile to the earliest/
latest scenario correspondingly.
Once a solution is developed, we assume the activities will start to be automated after the
cost of the solution falls below the level of wages for that activity. For this calculation, we took
into account both the evolution of solution costs and the evolution of wages.
124
Technical appendix
Solution cost evolution

Based on the capability requirements, solutions are classified into two categories, hardware
and software. If a solution requires sensory perception, fine motor skills and dexterity, gross
motor skills, or mobility, it is classified as a hardware solution. Otherwise, it will be classified
as a software solution. For a given solution, the initial cost is estimated as a percentage of
the highest hourly wage for the corresponding activity across all the countries we modeled.
We estimated the initial cost by looking at several examples of solutions developed using
different mixes of hardware and software. Based on our research, most of the software
solutions have relatively low initial costs as a percentage of the human labor cost. Some
solutions that require a combination of both software and hardware components have a
higher initial cost. To be conservative, we exclude certain solutions, whose advantages
are to be derived from only very specific scenarios or include non-economic benefits such
as increased quality and efficiency as well as decreased error rates. The range of initial
solution cost we model for hardware is 20 to 70percent of the highest hourly wage for the
corresponding activity in the world, and 0 to 20percent for software.
In our model, the solution costs decrease as technology advances, with hardware solution
costs declining by 16percent per year and software solution costs declining by 5.3percent
per year. We triangulate consumer price index and supplier surveys to estimate the
hardware solution cost reduction. For this, we use computer software and accessories
indexes to estimate software solution cost reduction, and, for, consumer price inflation, we
used consumer price index data for personal computers and peripheral equipment, and for
computer software and accessories from the US Bureau of Labor Statistics. For software
solution costs we used a survey of prices from the International Federation of Robotics.
Further work could be done to refine the estimation; however, given softwares low starting
cost, annual reduction has little impact on final automation results.
Wage evolution
We model the wage evolution for each country in two stages. From today until 2030, we
apply country-level growth estimates for all countries. The McKinsey Global Growth Model
provides wage data to 2030 for G20 countries and other major regions. We use best proxy
to estimate the remaining 26 countries not covered by the McKinsey model. As the data
are in local currency, we convert it into 2010 constant US dollars, by dividing nominal GDP
by corresponding country-level consumer prices (2010 base) and times the exchange rate
to the dollar (2010 base). We then calculated a compound annual growth rate (CAGR) for
each country. For 2030 onward, we group the 46 countries for which we have data into two
cohorts using a cutoff country-level annual wage, based on the wage distribution from 2016
to 2030. Countries within the same cohort grow at same rate. The cohort-level wage growth
rate is then estimated using the CAGR from 2016 to 2030 of the corresponding G20 country
within the cohort. The cutoff level also evolves at 2.66percent annually, which is the median
G20 CAGR between 2016 and 2030. We reclassify countries each year. As a country
advances into the next cohort, the appropriate growth rate will be applied. The detailed
proxy assignment and grouping is listed in ExhibitA5.
125
Exhibit A5
We approximate 26 countries with proper proxy till 2030 and classify them into two cohorts with group-level growth
rate for 2030 onward
McKinsey Global Growth Model (GGM) prediction and
proxy
GGM provides country-level growth rate for G19 + Spain
We use GGMs regional data to approximate the

other 26
Other Western Europe
Austria
Czech Republic
Netherlands
Norway
Sweden
Switzerland
Other sub-Saharan
Africa
Kenya
Nigeria
Saudi Arabia
Bahrain
Kuwait
Qatar
United Arab Emirates
Other Middle East and
North Africa
Egypt
Morocco
Oman
Other developed
East Asia and Pacific
Singapore
Other South Asia
Malaysia
Philippines
Thailand
Other Latin America
and Caribbean
Barbados
Chile
Colombia
Costa Rica
Peru
Other Eastern
Europe and Central
Asia
Poland
Other PIIGS
Greece
Cohort growth
24 high-income countries1 with 1.95% growth rate2
Australia
Austria
Bahrain
Canada
France
Germany
Greece
Italy
Japan
Kuwait
Netherlands
Norway
Qatar
Saudi Arabia
Singapore
South Korea
Spain
Sweden
Switzerland
United Arab
Emirates
United
Kingdom
United States
22 low-income countries1 with 3.12% growth rate2
Argentina
Barbados
Brazil
Chile
China
Colombia
Costa Rica
Czech Republic
Egypt
India
Indonesia
Kenya
Malaysia
Mexico
Morocco
Nigeria
Oman
Peru
Philippines
Poland
Russia
South Africa
Thailand
Turkey
Starting 2017, countries are regrouped each year

based on the evolving cut-off derived from
averaged country-level wage distribution
1 Adjusted UN standard based on wage distribution.

2 Growth rate is the median G19 compounded annual growth rate over 201630 for corresponding cohorts.
Adoption and deployment

Adoption can start once solutions are economically feasible, but several factors can still
hinder or enable both the timing and the pace of adoption. Solutions requiring different
technologies have different levels of ease. It takes time to integrate capabilities needed into
current technical platforms and combine them as an organic entity. Barriers also exist from
the organization side. Human talent and organization structures might act as bottlenecks to
implementation. Policies and law could also slow down or speed up technology innovation
and adoption. Finally, consumers might have varied preferences for automated solutions
due to perceived fears and emotional reactions, which could affect the adoption timing.
126
Technical appendix
Equations
To incorporate all these factors, we use the mathematics of the Bass diffusion model, a
well-known and widely used function in forecasting, especially for new products sales
forecasting and technology forecasting.
()
= +
1 ()
Where F(t) is the installed base fraction (that is, adoption of given technology or product)
and f(t) is the corresponding rate of change.
The function in our case also contains
two
key parameters:
= p parameter (inherent tendency

of consumers to adopt new technology), and q parameter (the tendency of consumers to

adopt based on peer adoption). Parameters are estimated through ordinary least square

q parameter
values from
meta-analyses can be
regression. In the absence
of data, p and
used if saturation value is known or can be guessed.
Exhibit A8
We then simulate two scenarios for known historic technology adoption curves (see
Exhibit21 in the main report). The technologies we used are stents, airbags, laparoscopic
surgery, MRI,
smartphones, TVs,
antilock braking systems, online air booking, cellphones,
19+
color TVs, SxEW leaching (copper), personal computers, electronic stability control,
19+

instrument landing systems,
dishwashers,
pacemakers.
The fitted values of parameters
and
119
p and q are consistent with other academic research. It takes about five years to
reach 50percent adoption in the earliest scenario and approximately 16years in the
latest scenario.
Equations
3. ECONOMIC MODELING
To better understand the economic implication of automation, we examine it at both the
global and country levels. Key assumptions are similar across the two levels. The global level
is essentially an aggregated version of major countries including the G19 and Nigeria.
Our analysis is based on three primary components at country level: projected number of
full-time equivalents (FTE), the number
() of full-time employees needed to maintain GDP per
capita and achieve GDP per capita projection,
= and
+automation
output.
1 ()
Projected number of full-time employees

The projected FTE data indicates the labor force evolution. For a given year in a given
country, it is calculated as:
=

19+
19+
Based on an empirical generalization from 218 technologies. See Fareena Sultan, John U. Farley, and Donald
R. Lehmann, Reflections on A meta-analysis of applications of diffusion models, Journal of Marketing
Research, volume 33, number 2, May 1996; time periods are annual.
119
127
We modeled different segments: ages 014, ages 1524, ages 2464 female, ages 2464
male, and ages 65-plus. For each segment we gathered data on:
Population: Our model uses the United Nations projection on population to 2065.
Labor participation rate: We use historical data from the International Labour
Organisation to 2012. We estimate future participation conservatively, taking the highest
rate between 2007 and 2012 and fixing it for the future for each segment. For the
participation rate of the 014 age segment, we assume it will remain the same as in 2012.
Unemployment rate: We again use data from the ILO and assume it will be at the longterm steady state average moving forward.
GDP per capita projection

From today to 2030, we use the GDP projection from the McKinsey Global Growth Model.
From 2030 to 2065, we made GDP projection with a cohort growth model. Based on the
distribution of GDP per capita as of today, we grouped countries into two cohorts and
calculated the cohort-level growth rate. Countries within same cohort grow at the same rate.
The cutoff of groupings also evolve every year at 1.58percent, which is the average CAGR of
all countries from 2016 to 2030.
The two cohorts are as follows:
Nine high-income countries, with a 1.08percent growth rate from 2030 onwards:
Australia, Canada, France, Germany, Italy, Japan, South Korea, United Kingdom, and
United States.
Eleven middle-income countries, with a 2.04percent growth rate from 2030: Argentina,
Brazil, China, India, Indonesia, Mexico, Nigeria, Russia, Saudi Arabia, South Africa,
and Turkey.
Number of FTEs needed to maintain GDP per capita or its projection

Assuming zero productivity growth, we calculate the required level of FTE to maintain both
the GDP per capita as of today and the projected GDP per capita. As to maintain current
GDP per capita, number of FTE needed is calculated by using current GDP per capita times
total population and divided by current productivity as of 2014. The number of FTE required
to maintain projected GDP per capita is calculated in a similar way. With projected GDP
per capita mentioned above, the result is then multiplied by total population and divided by
productivity as of 2014 to get the number of FTE needed.
Automation output
Using the projected FTE and estimated automation adoption, we calculate automation
output under different scenarios by multiplying them together. This enables us to gauge
automation impact in terms of human labor. To maintain consistency with other data
sources we leveraged, we made several additional assumptions. We consider only those job
activities that are available and well defined today. Also, we assume automation has a labor
substitution effect but no other performance gains, to be conservative. Finally, we assume
labor replaced by automation will rejoin the workforce at the same of productivity as today.
We also assume that additional output from automation will not decrease even if the total
number of FTEs declines as a result of demographic changes.
128
Technical appendix
Basic assumptions and calculations

remain
the same
at the
= global level. The only
difference is that we average key indicators such as global growth based on averaged GDP,

population, and participation rate as well as unemployment rate across G19 plus Nigeria.
We make an adjusted
calculation
for the number
of FTE
at a global
level. Given the nature
of FTE, it must be aggregated with adjustments based on corresponding country-level
productivity. We essentially recalculate the global level average productivity using countrylevel data as follows:
19+
19+
The global level averaged productivity is then divided by global-level GDP to calculate
corresponding number of FTE. Without this adjustment, the number of FTE needed to
maintain GDP growth will be inflated given difference on productivity across countries.
4. KEY PROXIES AND DATA SOURCES

All of the findings about automation derived from our model are backed up by the database
we built, which captures detailed information from country, sector, job, and specific activity
level, corresponding time spent, and wages for certain activities. At each industry, job, and
activity level, we also calculated the corresponding impact using wage and number of fulltime employees. ExhibitA6 shows the database structure and key calculations. We leverage
many data sources, both internal and external, however, some approximation was needed
since the data are not comprehensive.
Exhibit A6
Our approach to model design and estimating impact of automation
Database structure
Impact definition
Country
Industry
Impact by
job title
($ per year)
Number
of FTEs
Annual wage
($)
Impact by
activity
($ per year)
Number
of FTEs1
Time spent on
activity per year
(hours)
Impact by
industry
= Job title
Job title
Activity
Ease of automation
(high, medium, low)
Time per year spent
(hours)
Number
of FTEs
Average
hourly wage
($)
Annual wage
($)
Annual wage spent

on this activity ($)
Average wage
(hourly and annual)
Number of FTEs
1 FTE whose job titles include these detailed work activities.
129
To gain insights on the impacts of automation, it is important to look at occupations as well

as activity levels. However, time spent on each activity by occupation is not directly available.
The US Bureau of Labor Statistics 2014 O*Net database provides frequency scores for each
activity ranging from 1 (with a significant range of year level) to 7 (with a significant range of
hour level). The score is used to calculate corresponding time factors. We plot these scores
against the time factor, which is assigned for given integer frequency. Assuming there are
2,080 working hours per year, we scale to calculate time spent on each activity by job title.
ExhibitA7 illustrates our methodology using a cashier as an example.
Exhibit A7
We converted frequency to number of hours using best fitting curve and validated it with job description
based on our case studies
Cashier example
Time spent on activity
Hours per year
200
y = kx3.89
150
100
50
0.2
0
3.4
Activity repeats
annually or more
3.6
3.8
4.0
4.2
4.4
4.6
Frequency score
130
Technical appendix
4.8
5.0
5.2
5.4
5.6
5.8
6.0
6.2
6.4 6.6 6.8

Activity repeats
after every hour
Data mapping and approximations

To compensate for a lack of consistent categorizations and level granularity of labor data
across all countries, we made estimates by mapping, proxies, and approximations. For
major countries, we extracted high-level employment data for nine major occupations
across 19 sectors from Oxford Economic Forecasts. Indias data were used to estimate
the agriculture sector for Brazil. We then combined these data with our own international
employment database, which contains detailed occupation data for Australia, Brazil,
Canada, France, the United Kingdom, and the United States. For all other countries where
data were missing, we defined and assigned proxies to them, based on GDP per capita and
output per capita, sector mix, and education levels. The key proxy defining process contains
two parts, the proxy for 39 major countries and the proxy for selected African countries.
For the 39 major countries, we leveraged existing external and internal McKinsey &
Company resources and assigned different proxies to countries within each industry. Taking
agriculture, which has 28percent of the global employees, as an example, we use India
to approximate China and Indonesia, given the similarity in farm size and GDP per capita.
For smaller industries, we use a weighted combination between Brazil and the developed
countries for which we have detailed data.
131
Half a century ago, a US commission appointed by President Lyndon B. Johnson

examined automations effect on employment and the economy.
132
Technical appendix
BIBLIOGRAPHY
A
Acemoglu, Daron, and David H. Autor, Skills, tasks, and
technologies: Implications for employment and earnings,
in Handbook of Labor Economics, volume 4B, David Card
and Orley Ashenfelter, eds., Elsevier, 2011.
Arntz, Melanie, Terry Gregory, and Ulrich Zierahn, The risk
of automation for jobs in OECD countries: A comparative
analysis, OECD Social, Employment and Migration
working paper number 189, OECD, May 2016.
Autor, David, Why are there still so many jobs? The
history and future of workplace automation, Journal of
Economic Perspectives, volume 29, number 3, 2015.
Autor, David, Skills, education, and the rise of earnings
inequality among the other 99percent, Science, volume
344, issue 6186, May 2014.
Autor, David, and David Dorn, The growth of low-skill
service jobs and the polarization of the US labor market,
American Economic Review, volume 103, number 5,
August 2013.
Autor, David, Frank Levy, and Richard J. Murnane,
The skill content of recent technological change: An
empirical explanation, Quarterly Journal of Economics,
November 2003.
B
Bass, Frank M., Comments on A new product growth
model for consumer durables, Management Science,
volume 50, number 12, December 2004.
Brewster, Signe, A strong robot hand with a softer side,

MIT Technology Review, February 2016.
Brynjolfsson, Erik, and Andrew McAfee, The second
machine age: Work, progress, and prosperity in a time of
Bughin, Jacques, Taking the measure of the networked
enterprise, McKinsey Quarterly, October 2015.
C
Cannata, Giorgio, and Marco Maggiali, Design of a tactile
sensor for robot hands, in Sensors: Focus on tactile,
force and stress sensors, Jose Gerardo Rocha and
Senentxu Lanceros-Mendez, eds., InTech, 2008.
Capek, Karel, R.U.R. (Rossums universal robots), 1920.
The play is available at www.gutenberg.org.
Carr, Nicholas, The big switch: Rewiring the world, from
Edison to Google, W. W. Norton, 2008.
Chui, Michael, James Manyika, and Mehdi Miremadi,
The four fundamentals of workplace automation,
McKinsey Quarterly, November 2015.
Chui, Michael, James Manyika, and Mehdi Miremadi,
Where machines could replace humansand where
they cant (yet), McKinsey Quarterly, July 2016.
Crafts, Nicholas, Steam as a general purpose
technology: A growth accounting perspective, Economic
Journal, volume 114, issue 495, April 2004.
Bessen, James, Toil and technology, Finance &

Development, volume 52, number 1, March 2015.
Eisenstein, Elizabeth, The printing press as an agent of

change, Cambridge University Press, 1980.
Bivens, Josh, and Lawrence Mishal, Understanding

the historic divergence between productivity and a
typical workers pay: Why it matters and why its real,
Economic Policy Institute briefing paper number 406,
September 2015.
Evangelista, Rinaldo, Paolo Guerrieri, and Valentina

Meliciani, The economic impact of digital technologies in
Europe, Economics of Innovation and New Technology,
November 2014.
Blanco, Carlos A., Roco de la Fuente, Isabel Caballero,

and Mara Luz Rodrguez-Mndez, Beer discrimination
using a portable electronic tongue based on screenprinted electrodes, Journal of Food Engineering, volume
157, July 2015.
Bowles, Jeremy, The computerization of European jobs,
Bruegel, July 2014.
F
Ford, Martin, Rise of the robots: Technology and the
threat of a jobless future, Basic Books, 2015.
Forget, Evelyn L., The town with no poverty: A history of
the North American guaranteed annual income social
experiments, University of Manitoba, May 2008.
Frey, Carl Benedikt, and Michael A. Osborne, The

future of employment: How susceptible are jobs to
computerisation? Oxford Martin School, September
17, 2013.
Furman, Jason, Is this time different? The opportunities
and challenges of artificial intelligence, remarks at AI
Now: The Social and Economic Implications of Artificial
Intelligence Technologies in the Near Term conference in
New York, July 7, 2016.
Keynes, John Maynard, Economic possibilities for our

grandchildren, in Essays in Persuasion, Macmillan, 1933.
Kromann, Lene, and Anders Sorensen, Automation,
performance and international competition: Firm-level
comparisons of process innovation, Copenhagen
Business School, 2015.
Kurzweil, Robots master skills with deep learning
technique, May 22, 2015.
G
Gartner, Gartner predicts our digital future, October
6, 2015.
Gordon, Robert, The rise and fall of American growth:
The US standard of living since the Civil War, Princeton
University Press, 2016.
Lebergott, Stanley, Labor force and employment

18001960, in Output, employment, and productivity
in the United States after 1800, Dorothy S. Brady, ed.,
NBER, 1966.
Graetz, Georg, and Guy Michaels, Robots at work, Centre

for Economic Performance discussion paper 1335,
March 2015.
Marx, Karl, Grundrisse: Foundations of the critique

of political economy, 1858 (unpublished manuscript),
available at www.marxists.org/archive.
Gutherz, Cheryl, and Shira Baron, Why patients with

primary care physicians use the emergency department
for non-urgent care, Einstein Journal of Biology and
Medicine, volume 18, number 4, 2001.
McKinsey & Company, Impact dinternet sur lconomie

franaise: Comment internet transforme notre pays
(The internets impact on the French economy: How the
internet is transforming our country), March 2011.
McKinsey Global Institute in collaboration with McKinsey

Analytics, The age of analytics: Competing in a datadriven world, December 2016.
Haller, Stefanie A., and Iulia Siedschlag, Determinants

of ICT adoption: Evidence from firm-level data, Applied
Economics, volume 43, issue 26, 2011.
Harrysson, Martyn, Detlef Schoder, and Asin Tavakoli,
The evolution of social technologies, McKinsey
Quarterly, June 2016.
Hodson, Hal, Googles DeepMind AI can lip-read TV
shows better than a pro, New Scientist, November
21, 2016.
Hoe, Robert, A short history of the printing press and of
the improvements in printing machinery from the time of
Gutenberg up to the present day, 1902.
Holloway, Stephen, Straight and level: Practical airline
economics, Routledge, 1997.
Hornstein, Andreas, Per Krusell, and Giovanni L.
Violante, The effects of technical change on labor market
inequalities, Center for Economic Policy Studies working
paper number 113, July 2005.
K
Karshenas, Massoud and Paul Stoneman, Rank,
stock, order, and epidemic effects in the diffusion of
new process technologies: An empirical model, RAND
Journal of Economics, volume 24, issue 4, 1993.
134
Bibliography
McKinsey Global institute, The China effect on global

innovation, October 2015.
McKinsey Global Institute, Digital America: A tale of the
haves and the have-mores, December 2015.
McKinsey Global Institute, Digital Europe: Pushing the
frontier, capturing the benefits, June 2016.
McKinsey Global Institute, Digital globalization: The new
era of global flows, February 2016.
McKinsey Global Institute, Global growth: Can
productivity save the day in an aging world?
January 2015.
McKinsey Global Institute, Growth and renewal in the
United States: Retooling Americas economic engine,
February 2011.
McKinsey Global Institute, Indias technology opportunity:
Transforming lives, empowering people, December 2014.
McKinsey Global Institute, A labor market that works:
Connecting talent with opportunity in the digital age,
June 2015.
McKinsey Global Institute, Playing to win: The new global
competition for corporate profits, September 2015.
McKinsey Global Institute, Poorer than their parents? Flat

or falling incomes in advanced economies, July 2016.
McKinsey Global Institute, Saudi Arabia beyond oil:
The investment and productivity transformation,
December 2015.
McKinsey Global Institute, The world at work: Jobs, pay,
and skills for 3.5billion people, June 2012.
Mishel, Lawrence, and Kar-Fai Gee, Why arent workers
benefiting from labour productivity growth in the United
States? International Productivity Monitor, number 23,
spring 2012.
Monegain, Bernie, IBM Watson pinpoints rare form of
leukemia after doctors misdiagnosed patient, Healthcare
IT News, August 8, 2016.
N
Nemeroff, Evan, Survey shows average mortgage
closing time hits three-year low, National Mortgage
News, August 22, 2014.
Nichols, Austin, and Jesse Rothstein, The earned income
tax credit, NBER working paper number 21211, May 2015.
Nielsen, China maintains robust e-commerce growth,
March 2016.
O
OMahony, Mary, and Marcel P. Timmer, Output, input,
and productivity measures at the industry level: The EU
KLEMS database, Economic Journal, volume 119, issue
538, June 2009.
OReilly, Tim, Dont replace people. Augment them,
Medium.com, July 17, 2016.
Ott, Mack, The growing share of services in the US
economydegeneration or evolution? Federal Reserve
Bank of St. Louis Review, June/July 1987.
R
Reveley, Mary S., Jeffrey L. Briggs, Joni K. Evans, Sharon
M. Jones, Tolga Kurtoglu, Karen M. Leone, and Carl E.
Sandifer, Causal factors and adverse events of aviation
accidents and incidents related to integrated vehicle
health management, National Aeronautics and Space
Administration, March 2011.
S
Sklar, Julia, Making robots talk to each other, MIT
Technology Review, August 2015.
Smith, Aaron, and Janna Anderson, AI, robotics, and the
future of jobs, Pew Research Center, August 6, 2014.
Sultan, Fareena, John U. Farley, and Donald R. Lehmann,
Reflections on A meta-analysis of applications of
diffusion models, Journal of Marketing Research, volume
33, number 2, May 1996.
T
Tamersov, Acar, Elias Khalil, Bo Xie, Stephen L. Lenkey,
Bryan R. Routledge, Duen Horng Chau, and Shamkant
B. Navathe, Large-scale insider-trading analysis: Patterns
and discoveries, Georgia Institute of Technology,
August 2014.
U
US Department of Health, Education, and Welfare,
Technology and the American economy: Report of the
National Commission on Technology, Automation, and
Economic Progress, 1966.
US Executive Office of the President, Artificial intelligence,
automation, and the economy, December 2016.
V
Vecchione, Anthony, New automated packaging system
ensures patient safety, Drug Topics, November 22, 2004.
W
Wong, David H., Kenneth B. Yap, Bradley Turner, and
Nexhmi Rexha, Predicting the diffusion pattern of
Internet-based communication applications using Bass
model parameter estimates for email, Journal of Internet
Business, issue 9, 2011.
World Economic Forum, The future of jobs: Employment,
skills, and workforce strategy for the fourth Industrial
Revolution, January 2016.
Rodrik, Dani, Premature deindustrialization, NBER

working paper number 20935, February 2015.
Rogers, Everett M., Diffusion of innovations, The Free
Press, 1962.
Rosa, Hartmut, Social acceleration: A new theory of
modernity, Columbia University Press, 2013.
135
RELATED MGI RESEARCH
The age of analytics: Competing in a datadriven world (December 2016)

Big datas potential just keeps growing.
Taking full advantage means companies must
incorporate analytics into their strategic vision
and use it to make better, faster decisions.
Digital America: A tale of the haves and

have-mores (December 2015)
While the most advanced sectors, companies,
and individuals push the boundaries of
technology use, the US economy as a whole is
realizing only 18percent of its digital potential.
Independent work: Choice, necessity, and

the gig economy (October 2016)
The MGI report examines all the ways people
are earning income, as well as the challenges
independent work presents.
The Internet of things: Mapping the value

beyond the hype (June 2015)
If policymakers and businesses get it right,
linking the physical and digital worlds could
generate up to $1.1trillion a year in economic
value by 2025.
Poorer than their parents? Flat or

falling incomes in advanced economies
(July 2016)
The real incomes of about two-thirds of
households in 25 advanced economies were
flat or fell between 2005 and 2014. Without
action, this phenomenon could have corrosive
economic and social consequences.
A labor market that works: Connecting

talent with opportunity in the digital age
(June 2015)
Online talent platforms are increasingly
connecting people to the right work
opportunities. By 2025 they could add
$2.7trillion to global GDP, and begin to
ameliorate many of the persistent problems in
the worlds labor markets.
Digital Europe: Pushing the frontier,

capturing the benefits (June 2016)
Europe is operating below its digital potential.
Accelerating digitization could add trillions
of euros to economic growth in less than
a decade.
Global growth: Can productivity save the

day in an aging world? (January 2015)
Without action, global economic growth
will almost halve in the next 50years. This
MGI report offers a solution: a dramatic
improvement in productivity.
www.mckinsey.com/mgi
E-book versions of selected MGI reports are available at MGIs website,
Amazons Kindle bookstore, and Apples iBooks Store.
Download and listen to MGI podcasts on iTunes or at
www.mckinsey.com/mgi/publications/multimedia/
Cover image Oli Scarff/Staff/Getty Images News.
Contents page images: Go board Steve McAlister/Getty Images; high-tech
supermarket Birgit Reichert/dpa Picture-Alliance/Alamy Stock Photo;
man and robot Javier Larrea/age fotostock/Getty Images.

January 2017
Copyright McKinsey & Company
www.mckinsey.com/mgi
@McKinsey_MGI
McKinseyGlobalInstitute

A Future That Works Full Report MGI January 2017

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

A Future That Works Full Report MGI January 2017

Загружено:

Авторское право:

Доступные форматы

A FUTURE THAT WORKS:

Copyright McKinsey & Company 2017

A FUTURE THAT WORKS:

James Manyika | San Francisco

IngaMaurer, JanMischke, SebastianMontero, ThomasNetzer, RobinNuttall,

A restaurant in Yiwu, Zhejiang province of China.

Executive summary Page 1

1. The new frontier Page 23

A vision of future automation

4. Factors affecting the pace and extent of automation Page 65

A FUTURE THAT WORKS: AUTOMATION,

Technical, economic, and social factors will determine

A global force that will transform economies and the workforce

About 60% of occupations

Wages associated with technically

Labor associated with technically

Technical automation potential, %

France, Germany, Italy, Spain, and the United Kingdom.

Automation will boost global

Five factors affecting pace and

G19 plus Nigeria

Last 50 years Next 50 years

Scenarios around time spent on current work activities, %

McKinsey Global Institute

McKinsey Global Institute

GAUGING AUTOMATION POTENTIAL IN THE GLOBAL WORKPLACE TODAY

McKinsey Global Institute

A future that works: Automation, employment, and productivity

Process sales and

SOURCE: McKinsey Global Institute analysis

McKinsey Global Institute

Only a small percentage of occupations can be fully automated by adapting

Sewing machine operators,

Technical automation potential (%)

Stock clerks, travel agents,

Fashion designers, chief

<5% of occupations consist of

About 60% of occupations

A future that works: Automation, employment, and productivity

1 Managing and developing people.

The degree of automation potential varies considerably among sectors

McKinsey Global Institute

Ability to automate (%)

Health care and

SOURCE: US Bureau of Labor Statistics; McKinsey Global Institute analysis

McKinsey Global Institute

A future that works: Automation, employment, and productivity

Within sectors, too, there is considerable variation. In manufacturing, for example,

McKinsey Global Institute

4547 4749 4951

Labor associated with

McKinsey Global Institute

A future that works: Automation, employment, and productivity

FACTORS AFFECTING PACE AND EXTENT OF AUTOMATION

McKinsey Global Institute

of labor substitution (see BoxE2, Automation technologies could provide significant

Box E2. Automation technologies could provide significant performance

McKinsey Global Institute

A future that works: Automation, employment, and productivity

Automation adoption will take decades, across a wide range of

McKinsey Global Institute

Rest of the economy

A future that works: Automation, employment, and productivity

Modeling scenarios for the pace and extent of automation adoption